Merge pull request 'fix(e2e): increase liveness probe max_tokens from 4 to 32 (#2204 )' (#2209 ) from fix/2204-liveness-probe-max-tokens into main

Merge pull request 'ci(workflows): add bp-directive comments to 4 PR-emitting jobs (Task #77 )' (#2211 ) from fix/77-bp-directive-4-emitters into main
Merge pull request 'fix(e2e): satisfy MODEL_REQUIRED in peer-visibility staging gate (#2212 )' (#2214 ) from fix/peer-visibility-test-model-required-2212 into main
2026-06-04 06:23:57 +00:00 · 2026-06-04 06:20:04 +00:00 · 2026-06-04 06:19:19 +00:00 · 2026-06-04 06:19:17 +00:00 · 2026-06-03 22:57:20 -07:00 · 2026-06-04 05:51:17 +00:00
101 changed files with 7380 additions and 297 deletions
@@ -466,12 +466,40 @@ def fetch_log(target_url: str) -> str | None:

 def grep_fail_markers(log_text: str) -> list[str]:
    """Return up to 5 sample matching lines for any FAIL_PATTERNS hit.
-    Empty list = clean log."""
+    Empty list = clean log.
+
+    Heuristic: skip lines where the marker appears inside script source
+    (e.g. ``echo "::error::..."`` in a ``::group::Run`` block) rather
+    than actual execution output. The Gitea Actions log prints the raw
+    script before executing it; ``echo "::error::"`` lines in that
+    display are false positives.
+    """
    matches: list[str] = []
+    in_run_group = False
+    group_depth = 0
    for line in log_text.splitlines():
+        stripped = line.strip()
+        # Track Gitea Actions group markers so we can skip the
+        # ``::group::Run`` script-source display blocks.
+        if stripped.startswith("::group::Run"):
+            in_run_group = True
+            group_depth = 1
+            continue
+        if stripped == "::endgroup::":
+            if in_run_group:
+                in_run_group = False
+                group_depth = 0
+            continue
+        if in_run_group:
+            continue
        for pat in FAIL_PATTERNS:
            if pat in line:
-                # Truncate to keep error output bounded.
+                # Additional false-positive guard: ``echo "::error::"``
+                # is script source, not a runtime error emission.
+                if pat == "::error::":
+                    prefix = line[: line.index(pat)].strip()
+                    if prefix.endswith('echo') or prefix.endswith("echo '") or prefix.endswith('echo "'):
+                        break
                matches.append(line.strip()[:240])
                break
        if len(matches) >= 5:
@@ -364,6 +364,71 @@ def _api_json_optional(url: str, token: str) -> tuple[int, dict | None]:
        return exc.code, None


+def current_branch_head(env: dict[str, str]) -> str | None:
+    """Return the SHA at the tip of the deploy branch (main) per Gitea, or None.
+
+    Used to detect a *superseded* deploy job (see `superseded_by`). Fail-safe:
+    any read error / missing token returns None so the caller treats the job as
+    NOT superseded and the strict /buildinfo verify still runs. We never let an
+    unreadable head silently green a deploy.
+    """
+
+    token = env.get("GITEA_TOKEN", "").strip()
+    if not token:
+        return None
+    host = env.get("GITEA_HOST", "git.moleculesai.app")
+    repo = env.get("GITHUB_REPOSITORY", "molecule-ai/molecule-core")
+    # Deploy lane is on: push:main; the branch is always main here, but read it
+    # from the ref name when present so a future branch rename doesn't break us.
+    branch = env.get("GITHUB_REF_NAME", "").strip() or "main"
+    url = f"https://{host}/api/v1/repos/{repo}/branches/{quote(branch, safe='')}"
+    status, body = _api_json_optional(url, token)
+    if status != 200 or not isinstance(body, dict):
+        return None
+    commit = body.get("commit")
+    if isinstance(commit, dict):
+        head = commit.get("id") or commit.get("sha")
+        if isinstance(head, str) and head.strip():
+            return head.strip()
+    return None
+
+
+def superseded_by(env: dict[str, str]) -> str | None:
+    """Return the newer head SHA if THIS deploy job has been superseded, else None.
+
+    This workflow runs with no `concurrency:` (intentional — Gitea 1.22.6 cancels
+    queued runs, which is unacceptable for a prod deploy). When two main pushes
+    land close together, BOTH deploy-production jobs run. The newer push rolls the
+    fleet forward first; the OLDER job's strict /buildinfo verify then sees tenants
+    on the NEWER SHA and false-reds with "$slug is stale" — even though the fleet
+    is AHEAD, not behind. Git SHAs aren't ordered, so the verify can't tell ahead
+    from behind on its own (and /buildinfo exposes only git_sha, no build time).
+
+    Resolve it at the source of truth for ordering — the branch ref: if main's
+    current head is a DIFFERENT SHA than the one this job is deploying, a newer
+    commit has landed and this job is superseded; the newest job's verify is the
+    authoritative one. We return that head SHA so the caller can log it and exit
+    success early, skipping the strict-equality verify for this stale job.
+
+    Fail-safe: returns None (NOT superseded) when the head can't be read or equals
+    our SHA, so a genuinely-behind tenant under the LATEST deploy job still fails
+    the strict verify loudly. This never suppresses a real-stale signal — it only
+    excuses a job that is no longer the latest from asserting exact equality.
+    """
+
+    sha = env.get("GITHUB_SHA", "").strip()
+    if not sha:
+        return None
+    head = current_branch_head(env)
+    if not head:
+        return None
+    # SHA lengths can differ (short vs full); compare on the shorter prefix.
+    n = min(len(head), len(sha))
+    if head[:n].lower() == sha[:n].lower():
+        return None
+    return head
+
+
 def live_disable_flag(env: dict[str, str]) -> str:
    """Return a live disable value from Gitea variables when readable.

@@ -442,6 +507,14 @@ def main() -> int:
    sub.add_parser("plan", help="print production deploy plan as JSON")
    sub.add_parser("assert-enabled", help="fail if production deploy is currently disabled")
    sub.add_parser("wait-ci", help="block until required CI context is green")
+    sub.add_parser(
+        "check-superseded",
+        help=(
+            "exit 0 if a newer commit has landed on the deploy branch (this job "
+            "is superseded; prints the newer head SHA), exit 10 if this job is "
+            "still the latest"
+        ),
+    )
    rollout_parser = sub.add_parser("rollout", help="execute canary-first scoped production rollout")
    rollout_parser.add_argument("--plan", required=True, help="path to prod-auto-deploy plan JSON")
    rollout_parser.add_argument("--response", required=True, help="path to write aggregate response JSON")
@@ -457,6 +530,16 @@ def main() -> int:
        if args.command == "wait-ci":
            wait_for_ci_context(dict(os.environ))
            return 0
+        if args.command == "check-superseded":
+            newer = superseded_by(dict(os.environ))
+            if newer:
+                print(newer)
+                return 0
+            # Exit 10 (not 0, not 1): "this job is still the latest". The
+            # workflow treats only exit 0 as superseded; 10 means proceed to
+            # the strict verify. A non-zero code here is informational, not a
+            # failure — the workflow step swallows it.
+            return 10
        if args.command == "rollout":
            rollout_from_plan_file(args.plan, args.response, dict(os.environ))
            return 0
@@ -13,20 +13,26 @@ set -euo pipefail
 OWNER="${REPO%%/*}"
 NAME="${REPO##*/}"
 API="https://${GITEA_HOST}/api/v1"
-CONTEXT="${TEAM}-review / approved (pull_request)"
+# Branch-protection requires the (pull_request_target) context variant.
+# The refire path must post the EXACT BP-required name so the gate flips.
+CONTEXT="${TEAM}-review / approved (pull_request_target)"
 TARGET_URL="https://${GITEA_HOST}/${OWNER}/${NAME}/pulls/${PR_NUMBER}"

 authfile=$(mktemp)
+post_authfile=$(mktemp)
 prfile=$(mktemp)
 postfile=$(mktemp)
 # shellcheck disable=SC2329 # invoked by EXIT trap
 cleanup() {
-  rm -f "$authfile" "$prfile" "$postfile"
+  rm -f "$authfile" "$post_authfile" "$prfile" "$postfile"
 }
 trap cleanup EXIT

-chmod 600 "$authfile"
+chmod 600 "$authfile" "$post_authfile"
 printf 'header = "Authorization: token %s"\n' "$GITEA_TOKEN" > "$authfile"
+# STATUS_POST_TOKEN is narrow-scoped write:repository for explicit status POST.
+# Falls back to GITEA_TOKEN for backward compatibility (e.g. local test).
+printf 'header = "Authorization: token %s"\n' "${STATUS_POST_TOKEN:-$GITEA_TOKEN}" > "$post_authfile"

 code=$(curl -sS -o "$prfile" -w '%{http_code}' -K "$authfile" \
  "${API}/repos/${OWNER}/${NAME}/pulls/${PR_NUMBER}")
@@ -68,7 +74,7 @@ body=$(jq -nc \
  '{state:$state, context:$context, description:$description, target_url:$target_url}')

 code=$(curl -sS -o "$postfile" -w '%{http_code}' -X POST \
-  -K "$authfile" -H "Content-Type: application/json" \
+  -K "$post_authfile" -H "Content-Type: application/json" \
  -d "$body" \
  "${API}/repos/${OWNER}/${NAME}/statuses/${head_sha}")
 if [ "$code" != "200" ] && [ "$code" != "201" ]; then
@@ -895,6 +895,47 @@ def resolve_required_teams(item: dict[str, Any], high_risk: bool) -> list[str]:
    return list(item.get("required_teams") or [])


+# ---------------------------------------------------------------------------
+# CI status validation for testing-class AI acks (internal#760 CTO hardening)
+# ---------------------------------------------------------------------------
+
+# Slugs that require CI / all-required green before an AI ack is valid.
+_TESTING_CLASS_SLUGS = {"comprehensive-testing", "local-postgres-e2e", "staging-smoke"}
+
+# Human-only carve-out: these items can NEVER be acked by AI, regardless
+# of config drift. Any item in this set MUST NOT have ai_ack_eligible.
+# migration / schema are future-proofing — not yet in config items, but
+# the code guard rejects them proactively (CTO hardening, msg 1388c76f).
+_HUMAN_ONLY_SLUGS = {"root-cause", "no-backwards-compat", "migration", "schema"}
+
+
+def get_ci_status(client: GiteaClient, owner: str, repo: str, sha: str) -> str:
+    """Return the state of CI / all-required (pull_request) for `sha`.
+
+    Looks through the commit statuses and returns the state string
+    ("success", "failure", "pending", "error") or "missing" if the
+    context is not found. This prevents an AI agent from attesting
+    "tests pass" independently of the actual CI run.
+    """
+    code, data = client._req(  # noqa: SLF001
+        "GET", f"/repos/{owner}/{repo}/statuses/{sha}"
+    )
+    if code != 200:
+        return "unknown"
+    if not data or not isinstance(data, list):
+        return "missing"
+    # Gitea returns statuses newest-first. Find the latest for our context.
+    for status in data:
+        if status.get("context") == "CI / all-required (pull_request)":
+            return status.get("state", "unknown")
+    return "missing"
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+
 def main(argv: list[str] | None = None) -> int:
    p = argparse.ArgumentParser()
    p.add_argument("--owner", required=True)
@@ -988,6 +1029,9 @@ def main(argv: list[str] | None = None) -> int:
    # one membership lookup per team.
    team_member_cache: dict[tuple[str, int], bool | None] = {}

+    # Pre-resolve the ai-sop-ack team id once (None if the team does not exist).
+    ai_sop_ack_team_id = client.resolve_team_id(args.owner, "ai-sop-ack")
+
    def probe(slug: str, users: list[str]) -> list[str]:
        # `slug` may be either an items-key (compute_ack_state caller) OR
        # an n/a-gate key (compute_na_state caller). Previously this hard
@@ -1042,14 +1086,18 @@ def main(argv: list[str] | None = None) -> int:
                    file=sys.stderr,
                )
        approved: list[str] = []
+        rejected_ai_ineligible: list[str] = []
+        rejected_ci_not_green: list[str] = []
        for u in users:
+            # 1) Human required_teams membership check
+            in_human_team = False
            for tid in team_ids:
                cache_key = (u, tid)
                if cache_key not in team_member_cache:
                    team_member_cache[cache_key] = client.is_team_member(tid, u)
                result = team_member_cache[cache_key]
                if result is True:
-                    approved.append(u)
+                    in_human_team = True
                    break
                if result is None:
                    print(
@@ -1059,6 +1107,44 @@ def main(argv: list[str] | None = None) -> int:
                    )
                    # Treat as not-in-team for this user/team pair; loop
                    # may still find membership in another team.
+            if in_human_team:
+                approved.append(u)
+                continue
+
+            # 2) AI-sop-ack team membership check (only for items that allow it).
+            if slug in items_by_slug:
+                item = items_by_slug[slug]
+                # Defensive: human-only carve-out is enforced in code, not just
+                # config. Even if ai_ack_eligible were mistakenly added to a
+                # migration/schema item, the AI path is rejected here.
+                if slug in _HUMAN_ONLY_SLUGS:
+                    rejected_ai_ineligible.append(u)
+                    continue
+                if item.get("ai_ack_eligible") and ai_sop_ack_team_id is not None:
+                    cache_key = (u, ai_sop_ack_team_id)
+                    if cache_key not in team_member_cache:
+                        team_member_cache[cache_key] = client.is_team_member(
+                            ai_sop_ack_team_id, u
+                        )
+                    result = team_member_cache[cache_key]
+                    if result is True:
+                        # 2a) Testing-class items require real CI artifact evidence.
+                        if slug in _TESTING_CLASS_SLUGS:
+                            ci_state = get_ci_status(
+                                client, args.owner, args.repo, head_sha
+                            )
+                            if ci_state != "success":
+                                print(
+                                    f"::warning::AI ack for {slug} rejected: "
+                                    f"CI / all-required is {ci_state}, not success",
+                                    file=sys.stderr,
+                                )
+                                rejected_ci_not_green.append(u)
+                                continue
+                        approved.append(u)
+                        continue
+            # If we get here, user is not approved for this slug.
+            rejected_ai_ineligible.append(u)
        return approved

    ack_state = compute_ack_state(
@@ -21,6 +21,7 @@ Scenarios:
  T16_comments_generic_approval — reviews empty; comments have "APPROVED" by team member → exit 0
  T17_comments_no_approval   — reviews empty; comments have no approval keywords → exit 1
  T18_review_wrong_team_comment_right_team — review candidate 404s, comment candidate passes
+  T19_ai_sop_ack_approved — ai-sop-ack member APPROVED review → team probe 404 → exit 1

 Usage:
  FIXTURE_STATE_DIR=/tmp/x python3 _review_check_fixture.py 8080
@@ -116,6 +117,12 @@ class Handler(http.server.BaseHTTPRequestHandler):
                    {"state": "CHANGES_REQUESTED", "dismissed": False, "user": {"login": "bob"}, "commit_id": "abc1234"},
                    {"state": "APPROVED", "dismissed": False, "user": {"login": "core-devops"}, "commit_id": "abc1234"},
                ])
+            if sc == "T19_ai_sop_ack_approved":
+                # ai-sop-ack member submitted APPROVED review — must NOT count
+                # toward qa-review (team_id=20) or security-review (team_id=21).
+                return self._json(200, [
+                    {"state": "APPROVED", "dismissed": False, "user": {"login": "ai-reviewer"}, "commit_id": "abc1234"},
+                ])
            # Default: one non-author APPROVED
            return self._json(200, [
                {"state": "APPROVED", "dismissed": False, "user": {"login": "core-devops"}, "commit_id": "abc1234"},
@@ -157,6 +164,9 @@ class Handler(http.server.BaseHTTPRequestHandler):
                return self._empty(403)
            if sc == "T18_review_wrong_team_comment_right_team" and login == "core-devops":
                return self._empty(404)
+            if sc == "T19_ai_sop_ack_approved" and login == "ai-reviewer":
+                # ai-sop-ack member is NOT in qa (20) or security (21).
+                return self._empty(404)
            # T7_team_member: member
            return self._empty(204)

@@ -0,0 +1,244 @@
+"""Live-fire regression test for #2159 — gate auto-fire runtime verification.
+
+Static tests (test_gate_review_auto_fire.py) validate that the workflow YAML
+is structurally correct. This test validates the *runtime* path: submitting an
+APPROVED review to a PR whose head contains the current gate workflows causes
+Gitea Actions to queue the qa-review + security-review workflows and POST the
+branch-protection-required (pull_request_target) contexts within a reasonable
+window.
+
+Skipped when Gitea API credentials are not available. Intended for:
+  - manual developer verification
+  - CI jobs provisioned with a service-account token
+
+Environment:
+  GITEA_HOST            — default: git.moleculesai.app
+  GITEA_TOKEN           — token with read:repository + write:issues (for review POST)
+  REPO                  — default: molecule-ai/molecule-core
+  LIVEFIRE_PR_NUMBER    — optional; if omitted the test tries to find a
+                          suitable open PR automatically, or skips.
+  LIVEFIRE_TIMEOUT_SEC  — default: 120
+"""
+
+import base64
+import json
+import os
+import re
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import pytest
+
+import yaml
+
+GITEA_HOST = os.environ.get("GITEA_HOST", "git.moleculesai.app")
+GITEA_TOKEN = os.environ.get("GITEA_TOKEN", "")
+REPO = os.environ.get("REPO", "molecule-ai/molecule-core")
+LIVEFIRE_PR_NUMBER = os.environ.get("LIVEFIRE_PR_NUMBER", "")
+LIVEFIRE_TIMEOUT_SEC = int(os.environ.get("LIVEFIRE_TIMEOUT_SEC", "120"))
+
+REQUIRED_CONTEXTS = [
+    "qa-review / approved (pull_request_target)",
+    "security-review / approved (pull_request_target)",
+]
+
+skip_no_token = pytest.mark.skipif(
+    not GITEA_TOKEN,
+    reason="GITEA_TOKEN not set — live-fire test requires API credentials",
+)
+
+
+def _api(method: str, path: str, body: dict | None = None) -> tuple[int, dict]:
+    url = f"https://{GITEA_HOST}/api/v1{path}"
+    headers = {
+        "Authorization": f"token {GITEA_TOKEN}",
+        "Content-Type": "application/json",
+    }
+    data = json.dumps(body).encode() if body else None
+    req = urllib.request.Request(url, data=data, headers=headers, method=method)
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            raw = resp.read()
+            code = resp.status
+    except urllib.error.HTTPError as exc:
+        raw = exc.read()
+        code = exc.code
+    payload = json.loads(raw) if raw else {}
+    return code, payload
+
+
+def _get_pr(number: int) -> dict:
+    code, pr = _api("GET", f"/repos/{REPO}/pulls/{number}")
+    if code != 200:
+        pytest.fail(f"GET /pulls/{number} returned HTTP {code}: {pr}")
+    return pr
+
+
+def _list_open_prs() -> list[dict]:
+    code, prs = _api("GET", f"/repos/{REPO}/pulls?state=open&limit=50")
+    if code != 200:
+        pytest.fail(f"GET /pulls?state=open returned HTTP {code}: {prs}")
+    return prs
+
+
+def _pr_has_trigger_in_head(pr: dict) -> bool:
+    """Return True if the PR head contains pull_request_review in both workflows."""
+    head_sha = pr["head"]["sha"]
+    for wf_name in ("qa-review.yml", "security-review.yml"):
+        path = f"/repos/{REPO}/contents/.gitea/workflows/{wf_name}?ref={head_sha}"
+        code, payload = _api("GET", path)
+        if code != 200:
+            return False
+        raw = base64.b64decode(payload.get("content", "")).decode("utf-8")
+        wf = yaml.safe_load(raw)
+        on = wf.get(True) or wf.get("on") or {}
+        if isinstance(on, str):
+            if on != "pull_request_review":
+                return False
+        elif "pull_request_review" not in on:
+            return False
+    return True
+
+
+def _find_suitable_pr() -> dict:
+    if LIVEFIRE_PR_NUMBER:
+        pr = _get_pr(int(LIVEFIRE_PR_NUMBER))
+        if pr.get("state") != "open":
+            pytest.skip(f"PR {LIVEFIRE_PR_NUMBER} is not open")
+        return pr
+
+    prs = _list_open_prs()
+    for pr in prs:
+        if _pr_has_trigger_in_head(pr):
+            return pr
+    pytest.skip("No open PR found whose head contains the pull_request_review trigger")
+
+
+def _submit_approved_review(pr_number: int) -> dict:
+    code, review = _api(
+        "POST",
+        f"/repos/{REPO}/pulls/{pr_number}/reviews",
+        {"body": "Live-fire test APPROVED review", "event": "APPROVED"},
+    )
+    # 200 = created, 422 = review already exists (idempotent enough for our purposes)
+    if code not in (200, 201, 422):
+        pytest.fail(f"POST /pulls/{pr_number}/reviews returned HTTP {code}")
+    return review
+
+
+def _get_status_snapshot(sha: str) -> dict[str, dict]:
+    """Return mapping context -> {id, updated_at, target_url} for required contexts."""
+    code, statuses = _api("GET", f"/repos/{REPO}/statuses/{sha}?limit=100")
+    if code != 200:
+        return {}
+    result: dict[str, dict] = {}
+    for st in statuses:
+        ctx = st.get("context", "")
+        if ctx in REQUIRED_CONTEXTS:
+            result[ctx] = {
+                "id": st.get("id"),
+                "updated_at": st.get("updated_at", st.get("created_at", "")),
+                "target_url": st.get("target_url"),
+            }
+    return result
+
+
+def _extract_run_id(target_url: str | None) -> str | None:
+    """Extract the Actions run_id from a status target_url."""
+    if not target_url:
+        return None
+    m = re.search(r"/actions/runs/(\d+)", target_url)
+    return m.group(1) if m else None
+
+
+def _poll_fresh_statuses(
+    sha: str,
+    prior_snapshot: dict[str, dict],
+    timeout_sec: int = LIVEFIRE_TIMEOUT_SEC,
+) -> dict[str, dict]:
+    """Poll until required contexts appear fresh (newer timestamp, id, or run)."""
+    deadline = time.monotonic() + timeout_sec
+    found: dict[str, dict] = {}
+    while time.monotonic() < deadline:
+        code, statuses = _api("GET", f"/repos/{REPO}/statuses/{sha}?limit=100")
+        if code == 200:
+            for st in statuses:
+                ctx = st.get("context", "")
+                if ctx in REQUIRED_CONTEXTS:
+                    updated_at = st.get("updated_at", st.get("created_at", ""))
+                    status_id = st.get("id")
+                    target_url = st.get("target_url")
+                    prior = prior_snapshot.get(ctx, {})
+                    # Fresh if timestamp changed, id changed, or target_url changed.
+                    is_fresh = (
+                        ctx not in prior_snapshot
+                        or updated_at != prior.get("updated_at", "")
+                        or status_id != prior.get("id")
+                        or target_url != prior.get("target_url")
+                    )
+                    if is_fresh:
+                        found[ctx] = {
+                            "state": st.get("state", st.get("status", "")),
+                            "updated_at": updated_at,
+                            "id": status_id,
+                            "target_url": target_url,
+                        }
+        if all(ctx in found for ctx in REQUIRED_CONTEXTS):
+            return found
+        time.sleep(5)
+    return found
+
+
+@skip_no_token
+class TestGateAutoFireLive:
+    def test_auto_fire_posts_required_contexts(self):
+        """Submit APPROVED review; assert BP-required contexts appear fresh within timeout."""
+        pr = _find_suitable_pr()
+        pr_number = pr["number"]
+        head_sha = pr["head"]["sha"]
+
+        # Capture pre-existing status snapshot so we can prove FRESH contexts
+        # were posted after the review submission (not stale from a prior run).
+        prior_snapshot = _get_status_snapshot(head_sha)
+        prior_run_ids = {
+            _extract_run_id(s["target_url"])
+            for s in prior_snapshot.values()
+            if _extract_run_id(s["target_url"])
+        }
+
+        review = _submit_approved_review(pr_number)
+
+        found = _poll_fresh_statuses(head_sha, prior_snapshot)
+
+        missing = [ctx for ctx in REQUIRED_CONTEXTS if ctx not in found]
+        if missing:
+            pytest.fail(
+                f"After {LIVEFIRE_TIMEOUT_SEC}s, fresh contexts still missing: {missing}. "
+                f"Found: {found}. Prior snapshot: {prior_snapshot}. "
+                f"PR #{pr_number} head={head_sha}. "
+                f"This indicates the pull_request_review trigger did not fire at runtime."
+            )
+
+        # The contexts appeared fresh — that's the proof of auto-fire.
+        # We do NOT assert success vs failure; the evaluator decides that.
+        # The point of #2159 is that the workflows QUEUE and POST at all.
+        for ctx, info in found.items():
+            state = info["state"]
+            assert state in ("pending", "success", "failure"), (
+                f"Unexpected state {state!r} for {ctx}"
+            )
+
+            # CR2 Finding 1: prove a NEW workflow run was triggered, not just
+            # an in-place status update. Gitea 1.22.6 lacks REST /actions/runs/*
+            # endpoints, so we use the run_id embedded in the status target_url
+            # as a proxy for distinct run_id.
+            run_id = _extract_run_id(info.get("target_url"))
+            if run_id and run_id in prior_run_ids:
+                pytest.fail(
+                    f"Context {ctx!r} has target_url run_id {run_id} which existed "
+                    f"BEFORE the review was submitted. This means the status was "
+                    f"updated in-place by an existing run, not by a new workflow "
+                    f"run triggered from the pull_request_review event."
+                )
@@ -0,0 +1,168 @@
+"""Regression test #765 — gate auto-fire on real qa/security APPROVED review.
+
+Validates the structural configuration of qa-review.yml and security-review.yml
+so that a real team-member APPROVED review fires the workflow and POSTs the
+exact branch-protection-required context name. This is the test #2020's
+stale-context failure would have caught.
+"""
+
+from pathlib import Path
+
+import yaml
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+def load_workflow(name: str) -> dict:
+    with (ROOT / "workflows" / name).open() as f:
+        return yaml.safe_load(f)
+
+
+def _job_guard_string(workflow: dict) -> str:
+    """Return the raw job-level `if:` string for the single job."""
+    jobs = workflow["jobs"]
+    # Both qa-review and security-review have exactly one job named "approved".
+    job = jobs["approved"]
+    return str(job.get("if", ""))
+
+
+def _post_step(workflow: dict) -> dict:
+    """Return the explicit POST /statuses step from the job steps list."""
+    jobs = workflow["jobs"]
+    steps = jobs["approved"]["steps"]
+    for step in steps:
+        name = step.get("name", "")
+        if "Post required status context" in name:
+            return step
+    raise AssertionError("No explicit POST status step found")
+
+
+class TestQaReviewDirectTrigger:
+    def test_trigger_is_pull_request_review_submitted(self):
+        wf = load_workflow("qa-review.yml")
+        # PyYAML parses bare 'on' as boolean True.
+        on = wf[True]
+        assert "pull_request_review" in on, (
+            "qa-review must trigger on pull_request_review"
+        )
+        types = on["pull_request_review"].get("types", [])
+        assert "submitted" in types, (
+            "pull_request_review must include 'submitted' type"
+        )
+
+    def test_job_guard_requires_approved_state(self):
+        wf = load_workflow("qa-review.yml")
+        guard = _job_guard_string(wf)
+        assert "github.event.review.state == 'APPROVED'" in guard, (
+            "job guard must check review.state for 'APPROVED'"
+        )
+        assert "github.event.review.state == 'approved'" in guard, (
+            "job guard must check review.state for 'approved' (case fallback per #2135)"
+        )
+
+    def test_post_step_uses_status_post_token(self):
+        wf = load_workflow("qa-review.yml")
+        post = _post_step(wf)
+        env = post.get("env", {})
+        assert env.get("GITEA_TOKEN") == "${{ secrets.STATUS_POST_TOKEN }}", (
+            "POST step must use STATUS_POST_TOKEN for write-scoped status POST"
+        )
+
+    def test_post_step_context_name_exact(self):
+        """The context POSTed must byte-match the branch-protection requirement."""
+        wf = load_workflow("qa-review.yml")
+        post = _post_step(wf)
+        run = post.get("run", "")
+        assert '"qa-review / approved (pull_request_target)"' in run, (
+            "POST step must emit exact BP-required context name"
+        )
+
+
+class TestSecurityReviewDirectTrigger:
+    def test_trigger_is_pull_request_review_submitted(self):
+        wf = load_workflow("security-review.yml")
+        # PyYAML parses bare 'on' as boolean True.
+        on = wf[True]
+        assert "pull_request_review" in on, (
+            "security-review must trigger on pull_request_review"
+        )
+        types = on["pull_request_review"].get("types", [])
+        assert "submitted" in types, (
+            "pull_request_review must include 'submitted' type"
+        )
+
+    def test_job_guard_requires_approved_state(self):
+        wf = load_workflow("security-review.yml")
+        guard = _job_guard_string(wf)
+        assert "github.event.review.state == 'APPROVED'" in guard, (
+            "job guard must check review.state for 'APPROVED'"
+        )
+        assert "github.event.review.state == 'approved'" in guard, (
+            "job guard must check review.state for 'approved' (case fallback per #2135)"
+        )
+
+    def test_post_step_uses_status_post_token(self):
+        wf = load_workflow("security-review.yml")
+        post = _post_step(wf)
+        env = post.get("env", {})
+        assert env.get("GITEA_TOKEN") == "${{ secrets.STATUS_POST_TOKEN }}", (
+            "POST step must use STATUS_POST_TOKEN for write-scoped status POST"
+        )
+
+    def test_post_step_context_name_exact(self):
+        """The context POSTed must byte-match the branch-protection requirement."""
+        wf = load_workflow("security-review.yml")
+        post = _post_step(wf)
+        run = post.get("run", "")
+        assert '"security-review / approved (pull_request_target)"' in run, (
+            "POST step must emit exact BP-required context name"
+        )
+
+
+class TestRefireScriptContextName:
+    """review-refire-status.sh must emit the BP-required (pull_request_target) context."""
+
+    def test_refire_script_context_is_pull_request_target(self):
+        script = ROOT / "scripts" / "review-refire-status.sh"
+        content = script.read_text()
+        assert 'CONTEXT="${TEAM}-review / approved (pull_request_target)"' in content, (
+            "refire script CONTEXT must be the exact BP-required (pull_request_target) variant"
+        )
+        assert 'approved (pull_request)"' not in content, (
+            "refire script must NOT post bare (pull_request) context"
+        )
+
+
+class TestRefireTokenSeparation:
+    """The /qa-recheck + /security-recheck backstop must also use STATUS_POST_TOKEN."""
+
+    def _refire_step(self, workflow_name: str, step_name_keyword: str) -> dict:
+        wf = load_workflow(workflow_name)
+        jobs = wf["jobs"]
+        steps = jobs["review-refire"]["steps"]
+        for step in steps:
+            name = step.get("name", "")
+            if step_name_keyword in name:
+                return step
+        raise AssertionError(f"No refire step matching {step_name_keyword!r}")
+
+    def test_qa_refire_uses_status_post_token(self):
+        step = self._refire_step("sop-checklist.yml", "Refire qa-review")
+        env = step.get("env", {})
+        assert env.get("STATUS_POST_TOKEN") == "${{ secrets.STATUS_POST_TOKEN }}", (
+            "qa refire must receive STATUS_POST_TOKEN env var"
+        )
+        # Evaluator stays on read token
+        assert "SOP_TIER_CHECK_TOKEN" in env.get("GITEA_TOKEN", "") or "GITHUB_TOKEN" in env.get("GITEA_TOKEN", ""), (
+            "qa refire evaluator must stay on read-scoped token"
+        )
+
+    def test_security_refire_uses_status_post_token(self):
+        step = self._refire_step("sop-checklist.yml", "Refire security-review")
+        env = step.get("env", {})
+        assert env.get("STATUS_POST_TOKEN") == "${{ secrets.STATUS_POST_TOKEN }}", (
+            "security refire must receive STATUS_POST_TOKEN env var"
+        )
+        assert "SOP_TIER_CHECK_TOKEN" in env.get("GITEA_TOKEN", "") or "GITHUB_TOKEN" in env.get("GITEA_TOKEN", ""), (
+            "security refire evaluator must stay on read-scoped token"
+        )
@@ -0,0 +1,145 @@
+"""Stale-head diagnostic test for #2159.
+
+Deterministically reports whether a PR's HEAD contains the pull_request_review
+trigger in qa-review.yml and security-review.yml. If the trigger is absent,
+auto-fire on APPROVED review is impossible for that PR.
+
+This is used as a self-diagnostic for future stale-PR situations (PRs opened
+before #2157 merged, or branches cut from old bases).
+
+Environment:
+  GITEA_HOST  — default: git.moleculesai.app
+  GITEA_TOKEN — token with read:repository scope (optional; falls back to local files)
+  REPO        — default: molecule-ai/molecule-core
+  PR_NUMBER   — required when running against a real PR
+"""
+
+import base64
+import json
+import os
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import pytest
+
+import yaml
+
+GITEA_HOST = os.environ.get("GITEA_HOST", "git.moleculesai.app")
+GITEA_TOKEN = os.environ.get("GITEA_TOKEN", "")
+REPO = os.environ.get("REPO", "molecule-ai/molecule-core")
+PR_NUMBER = os.environ.get("PR_NUMBER", "")
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+def _api(method: str, path: str) -> tuple[int, dict]:
+    url = f"https://{GITEA_HOST}/api/v1{path}"
+    headers = {"Authorization": f"token {GITEA_TOKEN}"}
+    req = urllib.request.Request(url, headers=headers, method=method)
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            return resp.status, json.loads(resp.read())
+    except urllib.error.HTTPError as exc:
+        body = exc.read()
+        return exc.code, json.loads(body) if body else {}
+
+
+def _fetch_workflow_from_ref(workflow_name: str, ref: str) -> dict:
+    path = f"/repos/{REPO}/contents/.gitea/workflows/{workflow_name}?ref={ref}"
+    code, payload = _api("GET", path)
+    if code != 200:
+        pytest.fail(
+            f"GET {path} returned HTTP {code}: {payload}. "
+            f"Cannot determine whether PR head contains the trigger."
+        )
+    raw = base64.b64decode(payload.get("content", "")).decode("utf-8")
+    return yaml.safe_load(raw)
+
+
+def _fetch_workflow_local(workflow_name: str) -> dict:
+    p = ROOT / "workflows" / workflow_name
+    if not p.exists():
+        pytest.fail(f"Local workflow file not found: {p}")
+    return yaml.safe_load(p.read_text())
+
+
+def _has_pull_request_review_trigger(wf: dict) -> bool:
+    on = wf.get(True) or wf.get("on") or {}
+    if isinstance(on, list):
+        return "pull_request_review" in on
+    if isinstance(on, dict):
+        return "pull_request_review" in on
+    if isinstance(on, str):
+        return on == "pull_request_review"
+    return False
+
+
+def _diagnose_pr(pr_number: int) -> dict[str, bool]:
+    code, pr = _api("GET", f"/repos/{REPO}/pulls/{pr_number}")
+    if code != 200:
+        pytest.fail(f"GET /pulls/{pr_number} returned HTTP {code}: {pr}")
+
+    head_ref = pr["head"]["ref"]
+    head_sha = pr["head"]["sha"]
+
+    results: dict[str, bool] = {}
+    for wf_name in ("qa-review.yml", "security-review.yml"):
+        wf = _fetch_workflow_from_ref(wf_name, head_sha)
+        results[wf_name] = _has_pull_request_review_trigger(wf)
+
+    return {
+        "pr_number": pr_number,
+        "head_ref": head_ref,
+        "head_sha": head_sha,
+        "triggers": results,
+        "auto_fire_possible": all(results.values()),
+    }
+
+
+def _diagnose_local() -> dict[str, bool]:
+    results: dict[str, bool] = {}
+    for wf_name in ("qa-review.yml", "security-review.yml"):
+        wf = _fetch_workflow_local(wf_name)
+        results[wf_name] = _has_pull_request_review_trigger(wf)
+    return {
+        "pr_number": None,
+        "head_ref": "local-checkout",
+        "head_sha": None,
+        "triggers": results,
+        "auto_fire_possible": all(results.values()),
+    }
+
+
+class TestStaleHeadDiagnostic:
+    """Test deterministically reports 'auto-fire impossible for this PR' when
+    the PR head lacks the pull_request_review trigger.
+    """
+
+    def test_local_checkout_has_pull_request_review_trigger(self):
+        """Local files (the ones in this checkout) must contain the trigger.
+
+        This is the baseline: if the checkout itself is stale, every PR cut
+        from it will also be stale.
+        """
+        diag = _diagnose_local()
+        missing = [n for n, ok in diag["triggers"].items() if not ok]
+        if missing:
+            pytest.fail(
+                f"Local checkout is missing pull_request_review trigger in: {missing}. "
+                f"This branch cannot produce PRs that auto-fire."
+            )
+
+    @pytest.mark.skipif(not GITEA_TOKEN, reason="GITEA_TOKEN not set")
+    @pytest.mark.skipif(not PR_NUMBER, reason="PR_NUMBER not set")
+    def test_pr_head_has_pull_request_review_trigger(self):
+        """When PR_NUMBER is given, assert the PR head contains the trigger."""
+        diag = _diagnose_pr(int(PR_NUMBER))
+        if not diag["auto_fire_possible"]:
+            missing = [n for n, ok in diag["triggers"].items() if not ok]
+            pytest.fail(
+                f"Auto-fire impossible for PR #{diag['pr_number']}. "
+                f"Head ref={diag['head_ref']} sha={diag['head_sha']}. "
+                f"Missing trigger in: {missing}. "
+                f"This PR needs /qa-recheck + /security-recheck fallback, or a rebase onto current main."
+            )
@@ -486,3 +486,129 @@ def test_scoped_rollout_dry_run_does_not_assert_coverage():
        sleep=lambda _s: None,
    )
    assert aggregate["ok"] is True
+
+
+# --- Superseded-deploy guard (false-stale fix) -----------------------------
+#
+# Scenario this fixes: no `concurrency:` on the prod-deploy workflow means two
+# close main pushes run BOTH deploy-production jobs. eb31bcf (Fix A) and 286338
+# (Fix C) merge back-to-back; the 286338 job rolls the fleet to staging-2863380
+# first; the OLDER eb31bcf job's strict verify then sees tenants on 2863380 and
+# false-reds "stale" though the fleet is AHEAD. superseded_by detects that main's
+# head is no longer eb31bcf and lets the older job succeed without weakening the
+# behind-tenant signal for whichever job IS the latest.
+
+
+def test_superseded_by_returns_newer_head_when_main_moved_ahead(monkeypatch):
+    # eb31bcf job: main head is now 2863380 -> superseded, return the newer head.
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: "2863380fullhash")
+    newer = prod.superseded_by({"GITHUB_SHA": "eb31bcffullhash"})
+    assert newer == "2863380fullhash"
+
+
+def test_superseded_by_none_when_this_job_is_still_head(monkeypatch):
+    # 2863380 job (the latest): head == our SHA -> NOT superseded -> strict verify
+    # runs, so a genuinely-behind tenant still fails loudly.
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: "2863380fullhash")
+    assert prod.superseded_by({"GITHUB_SHA": "2863380fullhash"}) is None
+
+
+def test_superseded_by_matches_on_short_vs_full_sha_prefix(monkeypatch):
+    # GITHUB_SHA is full; Gitea may return a different-length id. Equal prefixes
+    # must NOT count as superseded (avoid false-skipping the real latest job).
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: "2863380")
+    assert prod.superseded_by({"GITHUB_SHA": "2863380fullhash"}) is None
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: "2863380FULLHASH")
+    assert prod.superseded_by({"GITHUB_SHA": "2863380fullhash"}) is None
+
+
+def test_superseded_by_fail_safe_returns_none_when_head_unreadable(monkeypatch):
+    # Fail-safe: unreadable head (no token / API error) must NOT be treated as
+    # superseded, so the strict verify still runs and never silently greens.
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: None)
+    assert prod.superseded_by({"GITHUB_SHA": "eb31bcffullhash"}) is None
+
+
+def test_superseded_by_none_without_github_sha(monkeypatch):
+    monkeypatch.setattr(prod, "current_branch_head", lambda _env: "2863380fullhash")
+    assert prod.superseded_by({}) is None
+
+
+def test_current_branch_head_parses_gitea_branch_commit_id(monkeypatch):
+    captured = {}
+
+    def fake_optional(url, _token):
+        captured["url"] = url
+        return 200, {"name": "main", "commit": {"id": "2863380fullhash"}}
+
+    monkeypatch.setattr(prod, "_api_json_optional", fake_optional)
+    head = prod.current_branch_head(
+        {"GITEA_TOKEN": "secret", "GITHUB_REPOSITORY": "molecule-ai/molecule-core"}
+    )
+    assert head == "2863380fullhash"
+    assert captured["url"].endswith("/repos/molecule-ai/molecule-core/branches/main")
+
+
+def test_current_branch_head_uses_ref_name_branch(monkeypatch):
+    captured = {}
+
+    def fake_optional(url, _token):
+        captured["url"] = url
+        return 200, {"commit": {"sha": "deadbeef"}}
+
+    monkeypatch.setattr(prod, "_api_json_optional", fake_optional)
+    head = prod.current_branch_head(
+        {"GITEA_TOKEN": "secret", "GITHUB_REF_NAME": "release"}
+    )
+    assert head == "deadbeef"
+    assert captured["url"].endswith("/branches/release")
+
+
+def test_current_branch_head_none_without_token():
+    assert prod.current_branch_head({}) is None
+
+
+def test_current_branch_head_none_on_non_200(monkeypatch):
+    monkeypatch.setattr(prod, "_api_json_optional", lambda _u, _t: (500, None))
+    assert prod.current_branch_head({"GITEA_TOKEN": "secret"}) is None
+
+
+# --- #2213: superseded check must fire BEFORE production side effects ----------
+#
+# Real incident shape: two main pushes land ~2 min apart. The OLDER deploy job
+# (GITHUB_SHA=7a72516, target staging-7a72516) started LATE — main head was
+# already 7f25373. The #2194 guard only protected the *verify* step, so the
+# older job still:
+#   1. rolled the canary (hongming) BACKWARD to staging-7a72516 (the #2213 red,
+#      seen as the newer job's verify reading hongming on the old SHA), then
+#   2. promoted :latest backward to the older image,
+# before finally skipping verify. The workflow now calls this same superseded
+# check BEFORE the redeploy + promote steps and gates both off when it fires.
+# These tests pin the contract that check-superseded relies on for the exact
+# incident shape.
+
+
+def test_superseded_by_fires_for_older_job_when_newer_already_head(monkeypatch):
+    # Older job (7a72516) re-checks the head just before rollout and finds the
+    # newer merge (7f25373) already owns main -> superseded -> skip side effects.
+    monkeypatch.setattr(
+        prod, "current_branch_head", lambda _env: "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+    )
+    newer = prod.superseded_by(
+        {"GITHUB_SHA": "7a72516f7e7ba1a710c4f393fef08be8d22e1866"}
+    )
+    assert newer == "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+
+
+def test_superseded_by_none_for_latest_job_so_it_still_rolls(monkeypatch):
+    # The newer job (7f25373) IS the head -> NOT superseded -> it proceeds to
+    # roll the fleet and verify, so a genuinely-behind tenant still fails loud.
+    monkeypatch.setattr(
+        prod, "current_branch_head", lambda _env: "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+    )
+    assert (
+        prod.superseded_by(
+            {"GITHUB_SHA": "7f25373309eca54a36f08c371ff783c3a47c3f8d"}
+        )
+        is None
+    )
@@ -205,6 +205,8 @@ chmod +x "$FIXTURE_DIR/bin/curl"
 # Helper: run the script with fixture environment
 run_review_check() {
  local scenario="$1"
+  local team="${2:-qa}"
+  local team_id="${3:-20}"
  echo "$scenario" >"$FIX_STATE_DIR/scenario"
  local out
  set +e
@@ -215,8 +217,8 @@ run_review_check() {
    REPO="molecule-ai/molecule-core" \
    PR_NUMBER="999" \
    DEFAULT_BRANCH="main" \
-    TEAM="qa" \
-    TEAM_ID="20" \
+    TEAM="$team" \
+    TEAM_ID="$team_id" \
    REVIEW_CHECK_DEBUG="0" \
    REVIEW_CHECK_STRICT="0" \
    bash "$SCRIPT" 2>&1
@@ -372,6 +374,25 @@ assert_eq "T18 exit code 0 (comment approval still considered)" "0" "$T18_RC"
 assert_contains "T18 comment candidate notice" "comment-based approval" "$T18_OUT"
 assert_contains "T18 comment approver accepted" "APPROVED by core-qa-agent" "$T18_OUT"

+# T19 — ai-sop-ack member APPROVED review must NOT count toward qa-review
+# or security-review (R1 hardening refinement, msg 1388c76f).
+echo
+echo "== T19 ai-sop-ack APPROVED review excluded from qa-review gate =="
+T19_OUT=$(run_review_check "T19_ai_sop_ack_approved" "qa" "20")
+T19_RC=$(cat "$FIX_STATE_DIR/last_rc")
+assert_eq "T19 exit code 1 (ai-sop-ack not in qa team)" "1" "$T19_RC"
+assert_contains "T19 ai-reviewer excluded from qa" "candidates: ai-reviewer" "$T19_OUT"
+assert_contains "T19 none are in qa team" "none are in team" "$T19_OUT"
+
+# T20 — same ai-sop-ack member must also be excluded from security-review gate.
+echo
+echo "== T20 ai-sop-ack APPROVED review excluded from security-review gate =="
+T20_OUT=$(run_review_check "T19_ai_sop_ack_approved" "security" "21")
+T20_RC=$(cat "$FIX_STATE_DIR/last_rc")
+assert_eq "T20 exit code 1 (ai-sop-ack not in security team)" "1" "$T20_RC"
+assert_contains "T20 ai-reviewer excluded from security" "candidates: ai-reviewer" "$T20_OUT"
+assert_contains "T20 none are in security team" "none are in team" "$T20_OUT"
+
 echo
 echo "------"
 echo "PASS=$PASS FAIL=$FAIL"
@@ -1003,3 +1003,299 @@ class TestComputeNaStateAcceptsGateNotInItems(unittest.TestCase):
            comments, "alice", na_gates, lambda *_: ["alice"]
        )
        self.assertFalse(na_state["security-review"]["declared"])
+
+
+# ---------------------------------------------------------------------------
+# internal#760 ceremony — ai-sop-ack team + ai_ack_eligible per-item flag
+# ---------------------------------------------------------------------------
+
+
+class TestAIAckEligibleConfig(unittest.TestCase):
+    """CTO-controlled allowlist (msg 1388c76f):
+      ai_ack_eligible: comprehensive-testing, local-postgres-e2e, staging-smoke,
+                       five-axis-review, memory-consulted
+      human-only:      root-cause, no-backwards-compat
+    """
+
+    def test_ai_ack_eligible_items(self):
+        cfg = sop.load_config(CONFIG_PATH)
+        items_by_slug = {it["slug"]: it for it in cfg["items"]}
+        eligible = {
+            "comprehensive-testing",
+            "local-postgres-e2e",
+            "staging-smoke",
+            "five-axis-review",
+            "memory-consulted",
+        }
+        for slug in eligible:
+            self.assertTrue(
+                items_by_slug[slug].get("ai_ack_eligible"),
+                f"{slug} must be ai_ack_eligible",
+            )
+
+    def test_human_only_items(self):
+        cfg = sop.load_config(CONFIG_PATH)
+        items_by_slug = {it["slug"]: it for it in cfg["items"]}
+        human_only = {"root-cause", "no-backwards-compat"}
+        for slug in human_only:
+            self.assertFalse(
+                items_by_slug[slug].get("ai_ack_eligible", False),
+                f"{slug} must NOT be ai_ack_eligible (human-only)",
+            )
+
+    def test_testing_class_slugs_constant(self):
+        """_TESTING_CLASS_SLUGS must match the three testing items."""
+        self.assertEqual(
+            sop._TESTING_CLASS_SLUGS,
+            {"comprehensive-testing", "local-postgres-e2e", "staging-smoke"},
+        )
+
+    def test_human_only_slugs_constant(self):
+        """_HUMAN_ONLY_SLUGS encodes the migration/schema carve-out.
+
+        If this set changes, the CTO must approve the widening.
+        """
+        self.assertEqual(
+            sop._HUMAN_ONLY_SLUGS,
+            {"root-cause", "no-backwards-compat", "migration", "schema"},
+        )
+
+    def test_human_only_invariant_enforced_in_code_and_config(self):
+        """Every config-present slug in _HUMAN_ONLY_SLUGS must be human-only.
+
+        This test fails if a migration/schema-class item accidentally
+        acquires ai_ack_eligible via config drift.  migration/schema are
+        future-proofing slugs not yet in the live config; they are checked
+        by the production probe closure but skipped here.
+        """
+        cfg = sop.load_config(CONFIG_PATH)
+        items_by_slug = {it["slug"]: it for it in cfg["items"]}
+        for slug in sop._HUMAN_ONLY_SLUGS:
+            if slug not in items_by_slug:
+                # Future-proofing slug (e.g. migration, schema) — not yet
+                # in config, but the code guard still rejects AI acks.
+                continue
+            self.assertFalse(
+                items_by_slug[slug].get("ai_ack_eligible", False),
+                f"{slug} is in _HUMAN_ONLY_SLUGS and must NEVER be ai_ack_eligible",
+            )
+
+
+class TestAIAckEligibilityProbe(unittest.TestCase):
+    """The probe closure in main() delegates to compute_ack_state.
+    We simulate the AI-ack path by injecting a probe that behaves like
+    the production probe (human team first, then ai-sop-ack fallback).
+    """
+
+    def setUp(self):
+        self.items = _items_by_slug()
+        self.aliases = _numeric_aliases()
+
+    def _probe_human_then_ai(self, human_users, ai_users):
+        """Return users in human_users immediately; users in ai_users only
+        if the item is ai_ack_eligible."""
+        def probe(slug, users):
+            item = self.items.get(slug, {})
+            approved = []
+            for u in users:
+                if u in human_users:
+                    approved.append(u)
+                elif u in ai_users and item.get("ai_ack_eligible"):
+                    approved.append(u)
+            return approved
+        return probe
+
+    def test_ai_ack_passes_for_eligible_item(self):
+        comments = [_comment("ai-bot", "/sop-ack five-axis-review")]
+        probe = self._probe_human_then_ai(human_users=set(), ai_users={"ai-bot"})
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["five-axis-review"]["ackers"], ["ai-bot"])
+
+    def test_ai_ack_rejected_for_human_only_item(self):
+        comments = [_comment("ai-bot", "/sop-ack root-cause")]
+        probe = self._probe_human_then_ai(human_users=set(), ai_users={"ai-bot"})
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["root-cause"]["ackers"], [])
+        self.assertIn("ai-bot", state["root-cause"]["rejected"]["not_in_team"])
+
+    def test_human_ack_still_works_for_ai_eligible_item(self):
+        comments = [_comment("bob", "/sop-ack comprehensive-testing")]
+        probe = self._probe_human_then_ai(human_users={"bob"}, ai_users=set())
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["comprehensive-testing"]["ackers"], ["bob"])
+
+    def test_ai_ack_rejected_for_testing_item_when_ci_red(self):
+        # Simulate the production probe that checks CI status for testing items.
+        # When CI is not green, ai-sop-ack member is rejected.
+        def probe(slug, users):
+            item = self.items.get(slug, {})
+            approved = []
+            for u in users:
+                if u == "ai-bot" and item.get("ai_ack_eligible"):
+                    # Testing items require CI green; simulate CI red.
+                    if slug in sop._TESTING_CLASS_SLUGS:
+                        continue  # rejected: CI not green
+                    approved.append(u)
+            return approved
+
+        comments = [_comment("ai-bot", "/sop-ack comprehensive-testing")]
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["comprehensive-testing"]["ackers"], [])
+
+    def test_ai_ack_passes_for_testing_item_when_ci_green(self):
+        # Simulate CI green → AI ack passes.
+        def probe(slug, users):
+            item = self.items.get(slug, {})
+            approved = []
+            for u in users:
+                if u == "ai-bot" and item.get("ai_ack_eligible"):
+                    if slug in sop._TESTING_CLASS_SLUGS:
+                        # CI is green → allow
+                        pass
+                    approved.append(u)
+            return approved
+
+        comments = [_comment("ai-bot", "/sop-ack comprehensive-testing")]
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["comprehensive-testing"]["ackers"], ["ai-bot"])
+
+
+class TestAIAckHumanOnlyMigrationSchema(unittest.TestCase):
+    """RC 8322: migration and schema items are human-only regardless of
+    any future config that might accidentally mark them ai_ack_eligible.
+
+    These slugs are not yet in the live config items list; the tests use
+    synthetic items so the production guard can be exercised directly.
+    """
+
+    def setUp(self):
+        # Synthetic items — if live config ever adds migration/schema,
+        # they MUST stay human-only. The probe below mirrors the actual
+        # production closure logic (human team first, then AI fallback
+        # with _HUMAN_ONLY_SLUGS guard).
+        self.items = {
+            "migration": {
+                "slug": "migration",
+                "ai_ack_eligible": True,
+                "required_teams": ["engineers"],
+            },
+            "schema": {
+                "slug": "schema",
+                "ai_ack_eligible": True,
+                "required_teams": ["engineers"],
+            },
+        }
+        self.aliases = {}
+
+    def _production_like_probe(self, human_users, ai_users):
+        """Return a probe that mirrors the production closure's guard."""
+
+        def probe(slug, users):
+            item = self.items.get(slug, {})
+            approved = []
+            for u in users:
+                if u in human_users:
+                    approved.append(u)
+                elif u in ai_users:
+                    # Production guard: _HUMAN_ONLY_SLUGS rejects AI acks
+                    # regardless of the ai_ack_eligible flag.
+                    if slug in sop._HUMAN_ONLY_SLUGS:
+                        continue
+                    if item.get("ai_ack_eligible"):
+                        approved.append(u)
+            return approved
+
+        return probe
+
+    def test_ai_ack_rejected_for_migration(self):
+        comments = [_comment("ai-bot", "/sop-ack migration")]
+        probe = self._production_like_probe(human_users=set(), ai_users={"ai-bot"})
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["migration"]["ackers"], [])
+        self.assertIn("ai-bot", state["migration"]["rejected"]["not_in_team"])
+
+    def test_ai_ack_rejected_for_schema(self):
+        comments = [_comment("ai-bot", "/sop-ack schema")]
+        probe = self._production_like_probe(human_users=set(), ai_users={"ai-bot"})
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["schema"]["ackers"], [])
+        self.assertIn("ai-bot", state["schema"]["rejected"]["not_in_team"])
+
+    def test_human_ack_still_works_for_migration(self):
+        # Human team member acking migration/schema is unaffected.
+        comments = [_comment("bob", "/sop-ack migration")]
+        probe = self._production_like_probe(human_users={"bob"}, ai_users=set())
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["migration"]["ackers"], ["bob"])
+
+    def test_human_ack_still_works_for_schema(self):
+        comments = [_comment("bob", "/sop-ack schema")]
+        probe = self._production_like_probe(human_users={"bob"}, ai_users=set())
+        state = sop.compute_ack_state(
+            comments, "alice", self.items, self.aliases, probe
+        )
+        self.assertEqual(state["schema"]["ackers"], ["bob"])
+
+
+class TestGetCIStatus(unittest.TestCase):
+    """Verify get_ci_status reads the correct context from commit statuses."""
+
+    def _client_with_statuses(self, statuses):
+        client = sop.GiteaClient("git.example.com", "tok")
+
+        def fake_req(method, path, body=None, ok_codes=(200, 201, 204)):
+            return 200, statuses
+
+        client._req = fake_req  # type: ignore[method-assign]
+        return client
+
+    def test_ci_green_returns_success(self):
+        client = self._client_with_statuses([
+            {"context": "CI / all-required (pull_request)", "state": "success"},
+        ])
+        self.assertEqual(
+            sop.get_ci_status(client, "o", "r", "sha1"), "success"
+        )
+
+    def test_ci_red_returns_failure(self):
+        client = self._client_with_statuses([
+            {"context": "CI / all-required (pull_request)", "state": "failure"},
+        ])
+        self.assertEqual(
+            sop.get_ci_status(client, "o", "r", "sha1"), "failure"
+        )
+
+    def test_missing_context_returns_missing(self):
+        client = self._client_with_statuses([
+            {"context": "some-other-context", "state": "success"},
+        ])
+        self.assertEqual(
+            sop.get_ci_status(client, "o", "r", "sha1"), "missing"
+        )
+
+    def test_api_error_returns_unknown(self):
+        client = sop.GiteaClient("git.example.com", "tok")
+
+        def fake_req(method, path, body=None, ok_codes=(200, 201, 204)):
+            return 500, {"error": "boom"}
+
+        client._req = fake_req  # type: ignore[method-assign]
+        self.assertEqual(
+            sop.get_ci_status(client, "o", "r", "sha1"), "unknown"
+        )
@@ -32,6 +32,26 @@
 # AUTHOR SELF-ACK IS FORBIDDEN regardless of which team contains them
 # — the gate script enforces commenter != PR author before checking
 # team membership.
+#
+# AI-SOP-ACK TEAM (internal#760 ceremony design, CTO-approved):
+#   The `ai-sop-ack` team contains AI agent identities that can ack
+#   SOP-checklist items ON BEHALF OF automated evidence.  An AI ack is
+#   only valid when:
+#     1. the item has `ai_ack_eligible: true`
+#     2. the item is NOT in the human-only carve-out (migration/schema)
+#     3. for testing-class items, CI / all-required (pull_request) is
+#        green on the current head SHA
+#
+#   AI acks NEVER count toward qa-review or security-review gates —
+#   those remain human-team-only (enforced by review-check.sh team
+#   probe against TEAM_ID 20/21).
+#
+#   INITIAL ai_ack_eligible allowlist (CTO-controlled, msg 1388c76f):
+#     comprehensive-testing, local-postgres-e2e, staging-smoke,
+#     five-axis-review, memory-consulted
+#   HUMAN-ONLY carve-out:
+#     root-cause, no-backwards-compat
+#   Any widening requires an explicit config change reviewed by CTO.

 version: 1

@@ -83,25 +103,31 @@ items:
    numeric_alias: 1
    pr_section_marker: "Comprehensive testing performed"
    required_teams: [qa, engineers]
+    ai_ack_eligible: true
    description: >-
      What was tested, how, edge cases covered. Ack from any qa-team
-      member (or engineers fallback while qa is small).
+      member (or engineers fallback while qa is small). AI ack valid
+      only when CI / all-required (pull_request) is green.

  - slug: local-postgres-e2e
    numeric_alias: 2
    pr_section_marker: "Local-postgres E2E run"
    required_teams: [engineers]
+    ai_ack_eligible: true
    description: >-
      Link to local CI artifact, or "N/A: pure-frontend change". Ack
      from any engineer who can verify the local DB test actually ran.
+      AI ack valid only when CI / all-required (pull_request) is green.

  - slug: staging-smoke
    numeric_alias: 3
    pr_section_marker: "Staging-smoke verified or pending"
    required_teams: [engineers]
+    ai_ack_eligible: true
    description: >-
      Link to canary run, or "scheduled post-merge". Ack from any
      engineer (core-devops/infra-sre are members of engineers team).
+      AI ack valid only when CI / all-required (pull_request) is green.

  - slug: root-cause
    numeric_alias: 4
@@ -120,6 +146,7 @@ items:
    numeric_alias: 5
    pr_section_marker: "Five-Axis review walked"
    required_teams: [engineers]
+    ai_ack_eligible: true
    description: >-
      Correctness / readability / architecture / security / performance.
      Ack from any non-author engineer.
@@ -140,6 +167,7 @@ items:
    numeric_alias: 7
    pr_section_marker: "Memory/saved-feedback consulted"
    required_teams: [engineers]
+    ai_ack_eligible: true
    description: >-
      List of feedback memories applicable to this change. Ack from
      any engineer who has the same memory access.
@@ -42,11 +42,9 @@ jobs:
  check:
    name: Migration version collision check
    runs-on: ubuntu-latest
-    # Phase 3 (RFC #219 §1): surface broken workflows without blocking
-    # the PR. Follow-up PR flips this off after surfaced defects are
-    # triaged.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # Phase 4 (RFC #219 §1): 22 days green since 2026-05-11 port.
+    # mc#1982 mask removed — no surfaced defects in this lane.
+    continue-on-error: false
    timeout-minutes: 5
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -96,6 +96,7 @@ env:
  GITHUB_SERVER_URL: https://git.moleculesai.app

 jobs:
+  # bp-exempt: advisory arm64 pilot, non-gating by design (internal#418).
  fast-checks:
    name: fast-checks
    # AND-set: only the Mac arm64 runner advertises macos-self-hosted.
@@ -123,8 +123,9 @@ jobs:
    # integration). See internal#512 for the class defect.
    runs-on: docker-host
    # Phase 3 (RFC #219 §1): surface broken workflows without blocking.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # mc#1982: mask removed. If regressions appear, root-fix the underlying
+    # test — do NOT renew the mask silently.
+    continue-on-error: false
    outputs:
      api: ${{ steps.decide.outputs.api }}
    steps:
@@ -160,8 +161,9 @@ jobs:
    # detect-changes for the full rationale.
    runs-on: docker-host
    # Phase 3 (RFC #219 §1): surface broken workflows without blocking.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # mc#1982: mask removed. If regressions appear, root-fix the underlying
+    # test — do NOT renew the mask silently.
+    continue-on-error: false
    timeout-minutes: 15
    env:
      # Unique per-run container names so concurrent runs on the host-
@@ -328,16 +330,40 @@ jobs:
      - name: Wait for /health
        if: needs.detect-changes.outputs.api == 'true'
        run: |
-          for i in $(seq 1 30); do
+          # Readiness signal: the platform binds /health only AFTER the full
+          # migration chain has been applied on cold start (it prints
+          # "Platform starting on :PORT" at that point). So a 200 from /health
+          # is the real "migrations done + server listening" signal.
+          #
+          # The migration chain grows every release, so a fixed ~30s budget is
+          # brittle by construction (it WILL be exceeded as migrations accrue).
+          # Use a generous wall-clock budget that comfortably exceeds
+          # cold-start + full-migration time, polling fast. This is robust to a
+          # growing chain WITHOUT masking a genuinely dead platform: if the
+          # background platform-server process has exited (e.g. a broken
+          # migration crashed it), we stop and fail loudly at once instead of
+          # waiting out the whole budget.
+          DEADLINE_SECS=180          # cold-start + full migration chain headroom
+          PLATFORM_PID="$(cat workspace-server/platform.pid 2>/dev/null || true)"
+          start=$(date +%s)
+          while :; do
            if curl -sf "$BASE/health" > /dev/null; then
-              echo "Platform up after ${i}s"
+              echo "Platform healthy after $(( $(date +%s) - start ))s"
              exit 0
            fi
+            # Fast-fail: if the platform process died, /health will never come.
+            if [ -n "$PLATFORM_PID" ] && ! kill -0 "$PLATFORM_PID" 2>/dev/null; then
+              echo "::error::platform-server (pid ${PLATFORM_PID}) exited before /health became reachable — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
+            if [ "$(( $(date +%s) - start ))" -ge "$DEADLINE_SECS" ]; then
+              echo "::error::Platform did not become healthy within ${DEADLINE_SECS}s — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
            sleep 1
          done
-          echo "::error::Platform did not become healthy in 30s"
-          cat workspace-server/platform.log || true
-          exit 1
      - name: Assert migrations applied
        if: needs.detect-changes.outputs.api == 'true'
        run: |
@@ -242,16 +242,36 @@ jobs:
      - name: Wait for /health
        if: needs.detect-changes.outputs.chat == 'true'
        run: |
-          for i in $(seq 1 30); do
+          # Readiness signal: the platform binds /health only AFTER the full
+          # migration chain has been applied on cold start (it prints
+          # "Platform starting on :PORT" at that point). So a 200 from /health
+          # is the real "migrations done + server listening" signal.
+          #
+          # The migration chain grows every release, so a fixed ~30s budget is
+          # brittle by construction. Use a generous wall-clock budget that
+          # comfortably exceeds cold-start + full-migration time, polling fast.
+          # Robust to a growing chain WITHOUT masking a dead platform: if the
+          # background platform-server process has exited, fail loudly at once.
+          DEADLINE_SECS=180          # cold-start + full migration chain headroom
+          PLATFORM_PID="$(cat workspace-server/platform.pid 2>/dev/null || true)"
+          start=$(date +%s)
+          while :; do
            if curl -sf "http://127.0.0.1:${PLATFORM_PORT}/health" > /dev/null; then
-              echo "Platform up after ${i}s"
+              echo "Platform healthy after $(( $(date +%s) - start ))s"
              exit 0
            fi
+            if [ -n "$PLATFORM_PID" ] && ! kill -0 "$PLATFORM_PID" 2>/dev/null; then
+              echo "::error::platform-server (pid ${PLATFORM_PID}) exited before /health became reachable — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
+            if [ "$(( $(date +%s) - start ))" -ge "$DEADLINE_SECS" ]; then
+              echo "::error::Platform did not become healthy within ${DEADLINE_SECS}s — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
            sleep 1
          done
-          echo "::error::Platform did not become healthy in 30s"
-          cat workspace-server/platform.log || true
-          exit 1

      - name: Install canvas dependencies
        if: needs.detect-changes.outputs.chat == 'true'
@@ -278,16 +298,38 @@ jobs:
          export NEXT_PUBLIC_WS_URL="ws://127.0.0.1:${PLATFORM_PORT}/ws"
          npx next dev --turbopack -p "${CANVAS_PORT}" > canvas.log 2>&1 &
          echo $! > canvas.pid
-          for i in $(seq 1 30); do
-            if curl -sf "http://localhost:${CANVAS_PORT}" > /dev/null 2>&1; then
-              echo "Canvas up after ${i}s"
-              exit 0
+          # Readiness must wait for the actual chat route to *compile*, not
+          # just for the dev server to bind the port. `next dev --turbopack`
+          # accepts the TCP connection well before it has compiled a route
+          # on first request, so a bare `curl /` can 200 (or hang) while the
+          # page the tests load is still building. We therefore probe the
+          # real route the specs navigate to (`/?m=chat`) and require a 2xx,
+          # which only happens once Turbopack has finished the first
+          # compile. The previous 30s budget was also too tight for a cold
+          # Turbopack first-compile on a loaded operator-host runner — the
+          # `Canvas did not start in 30s` flake. Raise to 120s (job
+          # timeout-minutes is 15, so this is comfortably bounded) and probe
+          # every 2s.
+          READY=""
+          for i in $(seq 1 60); do
+            # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+            # pollution of the captured status (lint-curl-status-capture.yml).
+            set +e
+            curl -s -o /dev/null -w '%{http_code}' "http://localhost:${CANVAS_PORT}/?m=chat" > /tmp/canvas-ready.code
+            set -e
+            CODE=$(cat /tmp/canvas-ready.code 2>/dev/null || echo "000")
+            if [ "$CODE" -ge 200 ] && [ "$CODE" -lt 400 ]; then
+              echo "Canvas (chat route compiled) up after ~$((i*2))s (HTTP ${CODE})"
+              READY=1
+              break
            fi
-            sleep 1
+            sleep 2
          done
-          echo "::error::Canvas did not start in 30s"
-          cat canvas.log || true
-          exit 1
+          if [ -z "$READY" ]; then
+            echo "::error::Canvas chat route did not compile in 120s (last HTTP ${CODE})"
+            cat canvas.log || true
+            exit 1
+          fi

      - name: Run Playwright E2E tests
        if: needs.detect-changes.outputs.chat == 'true'
@@ -130,13 +130,37 @@ jobs:
        run: |
          set -euo pipefail
          ./workspace-server/platform-server > workspace-server/platform.log 2>&1 &
-          echo $! > workspace-server/platform.pid
-          for i in $(seq 1 30); do
-            curl -sf "$BASE/health" >/dev/null && exit 0
+          PLATFORM_PID=$!
+          echo "$PLATFORM_PID" > workspace-server/platform.pid
+          # Readiness signal: the platform binds /health only AFTER the full
+          # migration chain has been applied on cold start (it prints
+          # "Platform starting on :PORT" at that point). So a 200 from /health
+          # is the real "migrations done + server listening" signal.
+          #
+          # The migration chain grows every release, so a fixed ~30s budget is
+          # brittle by construction. Use a generous wall-clock budget that
+          # comfortably exceeds cold-start + full-migration time, polling fast.
+          # Robust to a growing chain WITHOUT masking a dead platform: if the
+          # background platform-server process has exited, fail loudly at once.
+          DEADLINE_SECS=180          # cold-start + full migration chain headroom
+          start=$(date +%s)
+          while :; do
+            if curl -sf "$BASE/health" >/dev/null; then
+              echo "Platform healthy after $(( $(date +%s) - start ))s"
+              exit 0
+            fi
+            if ! kill -0 "$PLATFORM_PID" 2>/dev/null; then
+              echo "::error::platform-server (pid ${PLATFORM_PID}) exited before /health became reachable — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
+            if [ "$(( $(date +%s) - start ))" -ge "$DEADLINE_SECS" ]; then
+              echo "::error::Platform did not become healthy within ${DEADLINE_SECS}s — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
            sleep 1
          done
-          cat workspace-server/platform.log || true
-          exit 1

      - name: Run comprehensive E2E
        run: bash tests/e2e/test_comprehensive_e2e.sh
@@ -126,6 +126,7 @@ jobs:
  # push/dispatch/cron only (30+ min). This is NOT a fake-green mask of
  # the real assertion — it validates the driving script's bash syntax
  # and inline-python so a broken test script fails at PR time.
+  # bp-required: pending #1296 — PR emitter, not yet required (tracked in #1296).
  pr-validate:
    name: E2E Peer Visibility
    runs-on: ubuntu-latest
@@ -267,12 +268,36 @@ jobs:
          echo $! > platform.pid
      - name: Wait for /health
        run: |
-          for i in $(seq 1 30); do
-            curl -sf "$BASE/health" > /dev/null && { echo "Platform up after ${i}s"; exit 0; }
+          # Readiness signal: the platform binds /health only AFTER the full
+          # migration chain has been applied on cold start (it prints
+          # "Platform starting on :PORT" at that point). So a 200 from /health
+          # is the real "migrations done + server listening" signal.
+          #
+          # The migration chain grows every release, so a fixed ~30s budget is
+          # brittle by construction. Use a generous wall-clock budget that
+          # comfortably exceeds cold-start + full-migration time, polling fast.
+          # Robust to a growing chain WITHOUT masking a dead platform: if the
+          # background platform-server process has exited, fail loudly at once.
+          DEADLINE_SECS=180          # cold-start + full migration chain headroom
+          PLATFORM_PID="$(cat workspace-server/platform.pid 2>/dev/null || true)"
+          start=$(date +%s)
+          while :; do
+            if curl -sf "$BASE/health" > /dev/null; then
+              echo "Platform healthy after $(( $(date +%s) - start ))s"
+              exit 0
+            fi
+            if [ -n "$PLATFORM_PID" ] && ! kill -0 "$PLATFORM_PID" 2>/dev/null; then
+              echo "::error::platform-server (pid ${PLATFORM_PID}) exited before /health became reachable — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
+            if [ "$(( $(date +%s) - start ))" -ge "$DEADLINE_SECS" ]; then
+              echo "::error::Platform did not become healthy within ${DEADLINE_SECS}s — see log below"
+              cat workspace-server/platform.log || true
+              exit 1
+            fi
            sleep 1
          done
-          echo "::error::Platform did not become healthy in 30s"
-          cat workspace-server/platform.log || true; exit 1
      - name: Run LOCAL fresh-provision peer-visibility E2E (literal MCP list_peers)
        # HONEST gate — NO continue-on-error. The local backend uses
        # external-mode workspaces so this context tests the literal MCP
@@ -48,8 +48,10 @@ on:
      - 'workspace-server/internal/handlers/a2a_proxy.go'
      - 'workspace-server/internal/middleware/**'
      - 'workspace-server/internal/provisioner/**'
+      - 'workspace-server/internal/providers/providers.yaml'
      - 'tests/e2e/test_staging_full_saas.sh'
      - 'tests/e2e/lib/completion_assert.sh'
+      - 'tests/e2e/lib/model_slug.sh'
      - 'tests/e2e/lib/aws_leak_check.sh'
      - 'tests/e2e/test_aws_leak_check.sh'
      - '.gitea/workflows/e2e-staging-saas.yml'
@@ -61,8 +63,10 @@ on:
      - 'workspace-server/internal/handlers/a2a_proxy.go'
      - 'workspace-server/internal/middleware/**'
      - 'workspace-server/internal/provisioner/**'
+      - 'workspace-server/internal/providers/providers.yaml'
      - 'tests/e2e/test_staging_full_saas.sh'
      - 'tests/e2e/lib/completion_assert.sh'
+      - 'tests/e2e/lib/model_slug.sh'
      - 'tests/e2e/lib/aws_leak_check.sh'
      - 'tests/e2e/test_aws_leak_check.sh'
      - '.gitea/workflows/e2e-staging-saas.yml'
@@ -315,3 +319,148 @@ jobs:
            echo "::warning::saas teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
          fi
          exit 0
+
+  # ── PLATFORM-MANAGED BOOT REGRESSION (moonshot/kimi NOT_CONFIGURED) ──────────
+  #
+  # The REAL-boot complement to the deterministic unit suite
+  # (workspace_provision_platform_boot_test.go). Provisions a REAL staging
+  # claude-code workspace on the PLATFORM-managed path — provider=platform,
+  # model=moonshot/kimi-k2.6, NO tenant LLM key — and asserts it reaches
+  # status=online (NOT not_configured) and a completion returns 200, via the same
+  # online-wait + completion-assert the BYOK job uses.
+  #
+  # Why a SEPARATE job (not a matrix leg of e2e-staging-saas): the platform path
+  # injects NO secret and pins a different model, so its env block diverges from
+  # the BYOK job's. A dedicated job keeps each path's "verify key present" preflight
+  # honest (BYOK requires a key; platform requires its ABSENCE not to matter) and
+  # gives the regression its own named commit-status for branch protection.
+  #
+  # Add `E2E Staging Platform Boot` to branch protection after 3 consecutive
+  # green runs on main (de-flake window; this path shares the cp#245
+  # boot-timeout flake surface the BYOK job has, so it must prove stable before
+  # it can BLOCK — see the gate-making plan in the PR body).
+  # bp-required: pending #2187
+  e2e-staging-platform-boot:
+    name: E2E Staging Platform Boot
+    runs-on: ubuntu-latest
+    # Phase 3 (RFC #219 §1): surface without blocking until the de-flake window
+    # closes. mc#1982: do NOT renew this mask silently — the gate-making plan
+    # tracks the flip to false under #2187.
+    continue-on-error: true
+    timeout-minutes: 45
+    permissions:
+      contents: read
+
+    env:
+      MOLECULE_CP_URL: https://staging-api.moleculesai.app
+      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
+      # The regression combo: claude-code + platform-managed + moonshot/kimi-k2.6.
+      # NO E2E_*_API_KEY is set — platform-managed billing is owned by Molecule via
+      # the CP LLM proxy. The harness's E2E_LLM_PATH=platform branch sends empty
+      # secrets and pin-selects the platform model.
+      E2E_RUNTIME: claude-code
+      E2E_LLM_PATH: platform
+      # Smoke mode: a single parent workspace is enough to prove online +
+      # completion for the platform path (the A2A/delegation matrix is the BYOK
+      # job's job). Override E2E_DEFAULT_PLATFORM_MODEL via workflow_dispatch to
+      # exercise another platform model id.
+      E2E_MODE: smoke
+      E2E_RUN_ID: "platform-${{ github.run_id }}-${{ github.run_attempt }}"
+      E2E_KEEP_ORG: ${{ github.event.inputs.keep_org && '1' || '0' }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Verify admin token present
+        run: |
+          if [ -z "$MOLECULE_ADMIN_TOKEN" ]; then
+            echo "::error::CP_STAGING_ADMIN_API_TOKEN secret not set (Railway staging CP_ADMIN_API_TOKEN)"
+            exit 2
+          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var not set — EC2 leak verification cannot run"
+              exit 2
+            fi
+          done
+          echo "Admin token present ✓"
+
+      - name: Assert NO BYOK key leaks into the platform run
+        run: |
+          # The whole point of this job is the platform-managed path. A stray
+          # E2E_*_API_KEY in the runner env would (via the harness) still be
+          # skipped by the E2E_LLM_PATH=platform branch — but assert their
+          # absence loudly here so a future env edit can't silently convert this
+          # into a masked BYOK run that no longer exercises the regression.
+          for var in E2E_MINIMAX_API_KEY E2E_ANTHROPIC_API_KEY E2E_OPENAI_API_KEY; do
+            if [ -n "${!var:-}" ]; then
+              echo "::warning::$var is set in this platform-boot job's env — the harness ignores it on E2E_LLM_PATH=platform, but it should not be wired here."
+            fi
+          done
+          echo "Platform-managed path: no tenant LLM key required ✓"
+
+      - name: CP staging health preflight
+        run: |
+          code=$(curl -sS -o /dev/null -w "%{http_code}" --max-time 10 "$MOLECULE_CP_URL/health")
+          if [ "$code" != "200" ]; then
+            echo "::error::Staging CP unhealthy (got HTTP $code). Skipping — not a workspace bug."
+            exit 1
+          fi
+          echo "Staging CP healthy ✓"
+
+      - name: Run platform-managed boot E2E (online + completion)
+        id: e2e
+        run: bash tests/e2e/test_staging_full_saas.sh
+
+      - name: Teardown safety net (runs on cancel/failure)
+        if: always()
+        env:
+          ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+        run: |
+          set +e
+          orgs=$(curl -sS "$MOLECULE_CP_URL/cp/admin/orgs" \
+            -H "Authorization: Bearer $ADMIN_TOKEN" 2>/dev/null \
+            | python3 -c "
+          import json, sys, os, datetime
+          run_id = os.environ.get('GITHUB_RUN_ID', '')
+          d = json.load(sys.stdin)
+          today = datetime.date.today()
+          yesterday = today - datetime.timedelta(days=1)
+          dates = (today.strftime('%Y%m%d'), yesterday.strftime('%Y%m%d'))
+          # smoke mode slugs are e2e-smoke-YYYYMMDD-platform-<run_id>-...
+          if run_id:
+              prefixes = tuple(f'e2e-smoke-{d}-platform-{run_id}-' for d in dates)
+          else:
+              prefixes = tuple(f'e2e-smoke-{d}-platform-' for d in dates)
+          candidates = [o['slug'] for o in d.get('orgs', [])
+                        if any(o.get('slug','').startswith(p) for p in prefixes)
+                        and o.get('instance_status') not in ('purged',)]
+          print('\n'.join(candidates))
+          " 2>/dev/null)
+          leaks=()
+          for slug in $orgs; do
+            echo "Safety-net teardown: $slug"
+            set +e
+            curl -sS -o /tmp/plat-cleanup.out -w "%{http_code}" \
+              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+              -H "Authorization: Bearer $ADMIN_TOKEN" \
+              -H "Content-Type: application/json" \
+              -d "{\"confirm\":\"$slug\"}" >/tmp/plat-cleanup.code
+            set -e
+            code=$(cat /tmp/plat-cleanup.code 2>/dev/null || echo "000")
+            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+              echo "[teardown] deleted $slug (HTTP $code)"
+            else
+              echo "::warning::platform-boot teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/plat-cleanup.out 2>/dev/null)"
+              leaks+=("$slug")
+            fi
+          done
+          if [ ${#leaks[@]} -gt 0 ]; then
+            echo "::warning::platform-boot teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+          fi
+          exit 0
@@ -88,8 +88,9 @@ jobs:
    # surprises and keeps the routing rule discoverable in one place.
    runs-on: docker-host
    # mc#1982 Phase 3 (RFC §1): surface broken workflows without blocking.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # mc#1982: mask removed. If regressions appear, root-fix the underlying
+    # test — do NOT renew the mask silently.
+    continue-on-error: false
    outputs:
      handlers: ${{ steps.filter.outputs.handlers }}
    steps:
@@ -119,8 +120,9 @@ jobs:
    # exists). See detect-changes for the full routing rationale.
    runs-on: docker-host
    # mc#1982 Phase 3 (RFC §1): surface broken workflows without blocking.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # mc#1982: mask removed. If regressions appear, root-fix the underlying
+    # test — do NOT renew the mask silently.
+    continue-on-error: false
    env:
      # Unique name per run so concurrent jobs don't collide on the
      # bridge network. ${RUN_ID}-${RUN_ATTEMPT} is unique even across
@@ -251,6 +253,19 @@ jobs:
            echo "✓ $tbl table present"
          done

+      - if: needs.detect-changes.outputs.handlers == 'true'
+        name: Preflight — INTEGRATION_DB_URL must be present
+        run: |
+          # Belt-and-suspenders: if the postgres-start step failed to
+          # export INTEGRATION_DB_URL, fail loud BEFORE go test can
+          # t.Skip its way to a green build. Closes the workflow-level
+          # fail-open gap identified in PR #2166 blocker #2.
+          if [ -z "${INTEGRATION_DB_URL:-}" ]; then
+            echo "::error::INTEGRATION_DB_URL is empty — postgres-start step did not export the connection string"
+            exit 1
+          fi
+          echo "INTEGRATION_DB_URL is set"
+
      - if: needs.detect-changes.outputs.handlers == 'true'
        name: Run integration tests
        run: |
@@ -91,10 +91,10 @@ jobs:
    name: lint-bp-context-emit-match
    runs-on: ubuntu-latest
    timeout-minutes: 5
-    # Phase 3 (RFC #219 §1): surface drift without blocking. After 7
-    # clean scheduled runs on main, flip to false so a scheduled
-    # failure is a hard CI signal.
-    continue-on-error: true  # mc#1982 Phase 3 — flip to false after 7 clean main runs
+    # Phase 4 (RFC #219 §1): 22 days green since 2026-05-11 port,
+    # well past the 7-clean-run threshold. Scheduled failure is now
+    # a hard CI signal.
+    continue-on-error: false
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
@@ -48,11 +48,9 @@ jobs:
  scan:
    name: Scan workflows for curl status-capture pollution
    runs-on: ubuntu-latest
-    # Phase 3 (RFC #219 §1): surface broken workflows without blocking
-    # the PR. Follow-up PR flips this off after surfaced defects are
-    # triaged.
-    # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # Phase 4 (RFC #219 §1): 22 days green since 2026-05-11 port.
+    # mc#1982 mask removed — no surfaced defects in this lane.
+    continue-on-error: false
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - name: Find curl ... -w '%{http_code}' ... || echo "000" subshells
@@ -81,10 +81,10 @@ jobs:
    name: lint-required-context-exists-in-bp
    runs-on: ubuntu-latest
    timeout-minutes: 5
-    # Phase 3 (RFC #219 §1): surface the pattern without blocking PRs
-    # while the directive convention beds in. Follow-up flip to false
-    # after 7 clean days on main. mc#1982.
-    continue-on-error: true  # mc#1982 Phase 3 — flip to false after 7 clean main runs
+    # Phase 4 (RFC #219 §1): 22 days green since 2026-05-11 port,
+    # well past the 7-clean-day threshold. PR-time failure is now
+    # a hard CI signal.
+    continue-on-error: false
    steps:
      - name: Check out PR head with full history (need base SHA blobs)
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
@@ -49,37 +49,56 @@ jobs:
      GITHUB_SERVER_URL: https://git.moleculesai.app
    steps:
      - name: Identify runner
+        id: identify
+        continue-on-error: true
        run: |
          set -eu
          echo "arch=$(uname -m)"
          echo "kernel=$(uname -sr)"
          echo "shell=$BASH_VERSION"
          # Sanity: must actually be arm64. If amd64 sneaks in here,
-          # fail fast — that means the label routing is wrong.
+          # the job skips gracefully rather than hard-failing, because
+          # a mislabelled runner is an ops concern, not a code defect.
+          # Pilot lane must not make main red (#2146).
          case "$(uname -m)" in
-            aarch64|arm64) echo "arm64 confirmed" ;;
-            *) echo "ERROR: expected arm64, got $(uname -m)"; exit 1 ;;
+            aarch64|arm64)
+              echo "arm64 confirmed"
+              echo "arm64=true" >> "$GITHUB_OUTPUT"
+              ;;
+            *)
+              echo "ERROR: expected arm64, got $(uname -m) — label routing may be wrong"
+              echo "arm64=false" >> "$GITHUB_OUTPUT"
+              exit 1
+              ;;
          esac

      - name: Checkout
+        if: steps.identify.outputs.arm64 == 'true'
        uses: actions/checkout@v4
        with:
          fetch-depth: 1

      - name: Install shellcheck (arm64)
+        if: steps.identify.outputs.arm64 == 'true'
        continue-on-error: true
        run: |
          set -eu
          if command -v shellcheck >/dev/null 2>&1; then
            echo "shellcheck already present: $(shellcheck --version | head -1)"
          else
-            # Prefer apt if the runner base ships it; else download arm64 binary.
+            # Prefer apt if the runner base ships it; else download the
+            # correct platform binary (darwin vs linux).
            if command -v apt-get >/dev/null 2>&1; then
              sudo apt-get update -qq
              sudo apt-get install -y --no-install-recommends shellcheck
            else
              SC_VER=v0.10.0
-              curl -fsSL "https://github.com/koalaman/shellcheck/releases/download/${SC_VER}/shellcheck-${SC_VER}.linux.aarch64.tar.xz" \
+              if [ "$(uname -s)" = "Darwin" ]; then
+                SC_PKG="shellcheck-${SC_VER}.darwin.aarch64.tar.xz"
+              else
+                SC_PKG="shellcheck-${SC_VER}.linux.aarch64.tar.xz"
+              fi
+              curl -fsSL "https://github.com/koalaman/shellcheck/releases/download/${SC_VER}/${SC_PKG}" \
                | tar -xJf - --strip-components=1
              sudo mv shellcheck /usr/local/bin/
            fi
@@ -87,14 +106,15 @@ jobs:
          shellcheck --version | head -2

      - name: Run shellcheck on .gitea/scripts/*.sh
+        if: steps.identify.outputs.arm64 == 'true'
        continue-on-error: true
        run: |
          set -eu
          # Only the scripts we control under .gitea/scripts. Pilot
          # scope is intentionally narrow — broaden in a follow-up
          # once the lane is proven.
-          if ! command -v shellcheck >/dev/null 2>&1; then
-            echo "WARN: shellcheck binary not found — skipping (pilot mode)"
+          if ! command -v shellcheck >/dev/null 2>&1 || ! shellcheck --version >/dev/null 2>&1; then
+            echo "WARN: shellcheck not functional — skipping (pilot mode)"
            exit 0
          fi
          # NOTE: macOS ships Bash 3.2 (Apple license), no `mapfile`
@@ -16,14 +16,24 @@ name: publish-workspace-server-image
 #
 # Image tags produced:
 #   :staging-<sha> — per-commit digest, stable for canary verify
-#   :staging-latest — tracks most recent build on this branch
+#   :staging-latest — tracks most recent BUILD on this branch (set by the
+#                     build job, last-writer-wins, NOT prod-gated)
+#   :latest — tracks the most recent PROD-PROMOTED build. Re-pointed by the
+#             deploy-production job ONLY after green main CI + canary +
+#             fleet rollout + /buildinfo verification pass. So :latest ==
+#             "current prod image", never the raw build. (Added 2026-06-03
+#             after a stale :latest — last moved 2026-05-10 — reverted a
+#             production tenant on a no-arg redeploy.)
 #
 # Production auto-deploy:
 #   After both platform and tenant images are pushed, deploy-production waits
 #   for strict required push contexts on the same SHA to go green, then
 #   calls the production CP redeploy-fleet endpoint with target_tag=
-#   staging-<sha>. Set repo variable or secret PROD_AUTO_DEPLOY_DISABLED=true
-#   to stop production rollout while keeping image publishing enabled.
+#   staging-<sha>. On success (rollout + buildinfo verified) it re-points
+#   :latest to the same SHA. Set repo variable or secret
+#   PROD_AUTO_DEPLOY_DISABLED=true to stop production rollout while keeping
+#   image publishing enabled — in which case :latest is NOT advanced either
+#   (correct: an unpromoted build must not become :latest).
 #
 # Primary ECR target: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/*
 # Optional staging tenant mirror target:
@@ -252,7 +262,25 @@ jobs:
      PROD_AUTO_DEPLOY_BATCH_SIZE: ${{ vars.PROD_AUTO_DEPLOY_BATCH_SIZE || '3' }}
      PROD_AUTO_DEPLOY_DRY_RUN: ${{ vars.PROD_AUTO_DEPLOY_DRY_RUN || '' }}
      PROD_ALLOW_NON_PROD_CP_URL: ${{ vars.PROD_ALLOW_NON_PROD_CP_URL || '' }}
+      # #2213: per-tenant /buildinfo settle budget. A freshly-swapped tenant can
+      # keep serving the old image at the edge for a short drain window; the
+      # verify step polls each tenant up to this budget before declaring it stale.
+      PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS: ${{ vars.PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS || '240' }}
+      PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS: ${{ vars.PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS || '20' }}
    steps:
+      # The publish runner's default HOME (/home/hongming) is not writable, so
+      # git/docker credential saves fail (`Error saving credentials: mkdir
+      # /home/hongming: permission denied`) and halt the production rollout
+      # (#2193). Point HOME + DOCKER_CONFIG at the writable job temp dir —
+      # mirrors build-and-push's "Prepare writable Docker config" fix above.
+      - name: Prepare writable HOME + Docker config
+        run: |
+          set -euo pipefail
+          H="$RUNNER_TEMP/auto-deploy-home"
+          mkdir -p "$H/.docker"
+          echo "HOME=$H" >> "$GITHUB_ENV"
+          echo "DOCKER_CONFIG=$H/.docker" >> "$GITHUB_ENV"
+
      - name: Checkout
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

@@ -297,8 +325,50 @@ jobs:
          set -euo pipefail
          python3 .gitea/scripts/prod-auto-deploy.py wait-ci

-      - name: Call production CP redeploy-fleet
+      # Superseded-job guard — BEFORE any production side effect (#2213).
+      #
+      # This workflow has no `concurrency:` (see header: Gitea 1.22.6 cancels
+      # queued prod deploys). So two close main pushes run BOTH deploy-production
+      # jobs. The verify step already skips its strict /buildinfo check when this
+      # job is superseded (#2194) — but that guard was AFTER the redeploy and the
+      # :latest promote, so an OLDER job that started late still:
+      #   1. rolled the whole fleet BACKWARD to its older tag (canary hongming
+      #      was reverted from the newer SHA — the #2213 red), then
+      #   2. promoted :latest backward to the older image,
+      # and only THEN skipped verify and exited green. A superseded job must do
+      # NEITHER. We re-check the branch head here, immediately before the rollout,
+      # and skip every side effect when a newer commit already owns main.
+      #
+      # exit 0 + non-empty stdout => superseded (newer head printed); the redeploy
+      # and promote steps are gated off via this output. exit 10 => this job is
+      # still the latest, proceed to roll the fleet. Fail-safe: a head that can't
+      # be read returns NOT-superseded (exit 10), so a genuine deploy is never
+      # silently skipped. (Re-checked again at verify time to catch a newer job
+      # that lands DURING this rollout.)
+      - name: Check superseded before production side effects
+        id: supersede
        if: ${{ steps.plan.outputs.enabled == 'true' }}
+        run: |
+          set -euo pipefail
+          set +e
+          NEWER_HEAD="$(python3 .gitea/scripts/prod-auto-deploy.py check-superseded)"
+          SUPERSEDED_EXIT=$?
+          set -e
+          if [ "$SUPERSEDED_EXIT" -eq 0 ] && [ -n "$NEWER_HEAD" ]; then
+            echo "superseded=true" >> "$GITHUB_OUTPUT"
+            echo "::notice::Superseded before rollout: main head is now ${NEWER_HEAD:0:7} (this job deploys ${GITHUB_SHA:0:7}). Skipping redeploy + :latest promote so an older job never rolls the fleet backward."
+            {
+              echo "## Production auto-deploy skipped — superseded before rollout"
+              echo ""
+              echo "This deploy job's SHA \`${GITHUB_SHA:0:7}\` is no longer the head of \`main\` (now \`${NEWER_HEAD:0:7}\`)."
+              echo "A newer deploy job owns the fleet; rolling it backward to this older build would revert tenants and \`:latest\`. No side effects performed."
+            } >> "$GITHUB_STEP_SUMMARY"
+          else
+            echo "superseded=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Call production CP redeploy-fleet
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
        run: |
          set -euo pipefail
          python3 .gitea/scripts/prod-auto-deploy.py assert-enabled
@@ -357,18 +427,66 @@ jobs:
          fi

      - name: Verify reachable tenants report this SHA
-        if: ${{ steps.plan.outputs.enabled == 'true' }}
+        # Skip when superseded BEFORE rollout: the redeploy step did not run, so
+        # there is no redeploy-fleet response to verify against and the newer job
+        # owns verification (#2213). The in-step guard below still catches the
+        # case where a newer job lands DURING this job's rollout.
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
        env:
          TENANT_DOMAIN: moleculesai.app
        run: |
          set -euo pipefail
          RESP="$RUNNER_TEMP/prod-redeploy-response.json"
+
+          # Superseded-job guard. This workflow has no `concurrency:` (header
+          # explains why: Gitea 1.22.6 cancels queued prod deploys). So two
+          # close main pushes run BOTH deploy-production jobs. The newer one
+          # rolls the fleet to its (newer) build first; this older job's strict
+          # equality check below would then see tenants on the NEWER SHA and
+          # false-red "$slug is stale" even though the fleet is AHEAD, not
+          # behind (git SHAs aren't ordered; /buildinfo exposes only git_sha).
+          #
+          # If main's current head is no longer THIS job's SHA, a newer commit
+          # has landed and this deploy is superseded — the newest job's verify
+          # is authoritative. Skip strict verify and succeed. exit 0 => newer
+          # head printed (superseded); exit 10 => still the latest, proceed to
+          # the strict verify so a genuinely-behind tenant still fails loudly.
+          set +e
+          NEWER_HEAD="$(python3 .gitea/scripts/prod-auto-deploy.py check-superseded)"
+          SUPERSEDED_EXIT=$?
+          set -e
+          if [ "$SUPERSEDED_EXIT" -eq 0 ] && [ -n "$NEWER_HEAD" ]; then
+            echo "::notice::Superseded deploy: main head is now ${NEWER_HEAD:0:7} (this job deployed ${GITHUB_SHA:0:7}). The fleet is at or ahead of this build; the newer deploy job's verify is authoritative. Skipping strict SHA verify."
+            {
+              echo ""
+              echo "### Buildinfo verification skipped — superseded deploy"
+              echo ""
+              echo "This deploy job's SHA \`${GITHUB_SHA:0:7}\` is no longer the head of \`main\` (now \`${NEWER_HEAD:0:7}\`)."
+              echo "A newer deploy job is rolling the fleet forward; its verify is authoritative."
+            } >> "$GITHUB_STEP_SUMMARY"
+            exit 0
+          fi
+
          mapfile -t SLUGS < <(jq -r '.results[]? | .slug' "$RESP")
          if [ ${#SLUGS[@]} -eq 0 ]; then
            echo "::error::No tenants returned from redeploy-fleet; refusing to mark production deploy verified."
            exit 1
          fi

+          # Per-tenant settle/retry budget (#2213). A tenant whose container the
+          # CP just swapped can keep serving the OLD image at the edge for a short
+          # window while the old container drains — /buildinfo returns HTTP 200
+          # with the previous SHA, which `curl --retry` does NOT retry (it only
+          # retries connection/5xx failures, not a stale-but-200 body). Without a
+          # settle window a still-rolling tenant false-reds "stale" on the very
+          # first poll. So poll each tenant's /buildinfo until it reports the
+          # target SHA or the budget is exhausted; only THEN declare it stale or
+          # unreachable. This never masks a genuinely stuck tenant — a tenant that
+          # never reaches the target within the budget still fails loud (and the
+          # superseded-job revert class is already blocked before rollout above).
+          SETTLE_BUDGET_SECONDS="${PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS:-240}"
+          SETTLE_INTERVAL_SECONDS="${PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS:-20}"
+
          STALE_COUNT=0
          UNREACHABLE_COUNT=0
          UNHEALTHY_COUNT=0
@@ -380,18 +498,36 @@ jobs:
              continue
            fi
            url="https://${slug}.${TENANT_DOMAIN}/buildinfo"
-            body="$(curl -sS --max-time 30 --retry 3 --retry-delay 5 --retry-connrefused "$url" || true)"
-            actual="$(echo "$body" | jq -r '.git_sha // ""' 2>/dev/null || echo "")"
-            if [ -z "$actual" ]; then
-              echo "::error::$slug did not return /buildinfo after deploy."
-              UNREACHABLE_COUNT=$((UNREACHABLE_COUNT + 1))
-              continue
-            fi
-            if [ "$actual" != "$GITHUB_SHA" ]; then
-              echo "::error::$slug is stale: actual=${actual:0:7}, expected=${GITHUB_SHA:0:7}"
-              STALE_COUNT=$((STALE_COUNT + 1))
-            else
+            deadline=$(( $(date +%s) + SETTLE_BUDGET_SECONDS ))
+            actual=""
+            last_actual=""
+            on_target=false
+            while :; do
+              body="$(curl -sS --max-time 30 --retry 3 --retry-delay 5 --retry-connrefused "$url" || true)"
+              actual="$(echo "$body" | jq -r '.git_sha // ""' 2>/dev/null || echo "")"
+              [ -n "$actual" ] && last_actual="$actual"
+              if [ "$actual" = "$GITHUB_SHA" ]; then
+                on_target=true
+                break
+              fi
+              now=$(date +%s)
+              if [ "$now" -ge "$deadline" ]; then
+                break
+              fi
+              # Still rolling (stale 200) or transiently unreachable — wait and
+              # re-poll within the settle budget rather than failing on first read.
+              remaining=$(( deadline - now ))
+              echo "$slug: waiting for target SHA (have '${actual:0:7}', want ${GITHUB_SHA:0:7}; ${remaining}s left)"
+              sleep "$SETTLE_INTERVAL_SECONDS"
+            done
+            if [ "$on_target" = true ]; then
              echo "$slug: ${actual:0:7}"
+            elif [ -z "$last_actual" ]; then
+              echo "::error::$slug did not return /buildinfo after deploy (waited ${SETTLE_BUDGET_SECONDS}s)."
+              UNREACHABLE_COUNT=$((UNREACHABLE_COUNT + 1))
+            else
+              echo "::error::$slug is stale: actual=${last_actual:0:7}, expected=${GITHUB_SHA:0:7} (waited ${SETTLE_BUDGET_SECONDS}s)"
+              STALE_COUNT=$((STALE_COUNT + 1))
            fi
          done

@@ -409,3 +545,69 @@ jobs:
          if [ "$STALE_COUNT" -gt 0 ] || [ "$UNHEALTHY_COUNT" -gt 0 ] || [ "$UNREACHABLE_COUNT" -gt 0 ]; then
            exit 1
          fi
+
+      # Re-point :latest to the just-promoted image — ONLY after the
+      # production rollout + buildinfo verification above have passed.
+      #
+      # WHY HERE (promote point), not at build time:
+      #   The platform-tenant ECR `:latest` tag was last moved 2026-05-10
+      #   and went 3.5 weeks stale because the build step only pushes
+      #   :staging-<sha> + :staging-latest and never re-points :latest. A
+      #   no-arg POST /cp/admin/tenants/:slug/redeploy (whose default tag
+      #   fell through to "latest") then pulled the 3.5-week-old image and
+      #   REVERTED the tenant (incident: molecule-adk-demo, 2026-06-03).
+      #
+      #   The defense-in-depth half of this fix changes that redeploy
+      #   default to :staging-latest, but :latest itself must also be
+      #   kept meaningful. We make :latest track the PROD-BLESSED build,
+      #   not the raw build: by living at the end of deploy-production —
+      #   after `wait-ci` (green main CI), the canary-first batched fleet
+      #   rollout, AND the /buildinfo SHA verification — :latest only ever
+      #   advances to a SHA that is actually green and confirmed running
+      #   across the live fleet. So `:latest` == "current prod image",
+      #   and any consumer that pulls :latest (legacy callers, manual
+      #   `docker pull`, a redeploy that somehow still resolves "latest")
+      #   gets the blessed image instead of whatever happened to build.
+      #
+      #   Re-tag is digest-level (imagetools create), so no rebuild and
+      #   :latest is byte-identical to :staging-<sha> for this commit.
+      # Gate on supersede: a superseded older job must NOT move :latest backward
+      # to its older image (#2213 — 275383 promoted :latest → the older
+      # staging-7a72516 after a newer job had already shipped). :latest must only
+      # ever advance under the job that owns main's head.
+      - name: Promote :latest to the verified prod image
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
+        env:
+          TENANT_IMAGE_NAME: ${{ env.TENANT_IMAGE_NAME }}
+          STAGING_TENANT_IMAGE_NAME: ${{ env.STAGING_TENANT_IMAGE_NAME }}
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_DEFAULT_REGION: us-east-2
+        run: |
+          set -euo pipefail
+          SHA_TAG="staging-${GITHUB_SHA::7}"
+          PROD_ECR_REGISTRY="${TENANT_IMAGE_NAME%%/*}"
+          STAGING_ECR_REGISTRY="${STAGING_TENANT_IMAGE_NAME%%/*}"
+
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${PROD_ECR_REGISTRY}"
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${STAGING_ECR_REGISTRY}"
+
+          # imagetools create copies the source manifest to the new tag by
+          # digest (no pull/rebuild). :latest now points at the exact image
+          # that just passed the prod gate.
+          docker buildx imagetools create \
+            --tag "${TENANT_IMAGE_NAME}:latest" \
+            "${TENANT_IMAGE_NAME}:${SHA_TAG}"
+          docker buildx imagetools create \
+            --tag "${STAGING_TENANT_IMAGE_NAME}:latest" \
+            "${STAGING_TENANT_IMAGE_NAME}:${SHA_TAG}"
+
+          {
+            echo ""
+            echo "### :latest promoted"
+            echo ""
+            echo "Re-pointed \`platform-tenant:latest\` → \`${SHA_TAG}\` (prod + staging ECR)."
+            echo ":latest now tracks the prod-blessed, fleet-verified image."
+          } >> "$GITHUB_STEP_SUMMARY"
@@ -9,10 +9,22 @@
 #   Triggers on:
 #     - `pull_request_target`: opened, synchronize, reopened
 #         → initial status posts when PR opens / re-pushes
-#     - comment refires are handled by `review-refire-comments.yml`
-#         → a single issue_comment dispatcher prevents every SOP/review
-#           comment from enqueueing separate qa/security/tier jobs on
-#           Gitea 1.22.6 before job-level `if:` can skip them.
+#     - `pull_request_review` types: [submitted]
+#         → re-evaluate when a team member submits an APPROVE review so
+#           the gate flips immediately (no wait for the next push or
+#           slash-command). Verified live: sop-tier-check.yml uses this
+#           same event and provably fires (produces
+#           `sop-tier-check / tier-check (pull_request_review)` contexts).
+#           The job-level `if:` guard checks
+#           `github.event.review.state == 'APPROVED' || 'approved'` so
+#           only APPROVE reviews run the evaluator; COMMENT and
+#           REQUEST_CHANGES are skipped at the job level.
+#           Branch-protection requires the `(pull_request_target)`
+#           context variant, so the review-event path EXPLICITLY POSTS
+#           the required context via the API. Trust boundary preserved
+#           (BASE ref, no PR-head).
+#     - comment refires are handled by `sop-checklist.yml` review-refire job
+#         → `/qa-recheck` slash-command re-evaluates this gate.
 #   Workflow name = `qa-review` ; job name = `approved`.
 #   The job's own pass/fail conclusion publishes the status context
 #   `qa-review / approved (<event>)` — NO `POST /statuses` call → NO
@@ -85,21 +97,26 @@ name: qa-review
 on:
  pull_request_target:
    types: [opened, synchronize, reopened]
+  pull_request_review:
+    types: [submitted]

 permissions:
  contents: read
  pull-requests: read
-  secrets: read
+  statuses: write

 jobs:
  # bp-exempt: PR review bot signal; required merge state is enforced by CI / all-required.
  approved:
    # Gate the job:
    #   - On pull_request_target events: always run.
-    # Comment-triggered refires live in review-refire-comments.yml. Keeping
-    # this workflow PR-only avoids comment-triggered queue storms.
+    #   - On pull_request_review_approved events: run so the gate flips
+    #     immediately when a team member submits an APPROVE review.
+    # Comment-triggered refires live in sop-checklist.yml review-refire job.
    if: |
-      github.event_name == 'pull_request_target'
+      github.event_name == 'pull_request_target' ||
+      (github.event_name == 'pull_request_review' &&
+       (github.event.review.state == 'APPROVED' || github.event.review.state == 'approved'))
    runs-on: ubuntu-latest
    steps:
      - name: Privilege check (A1.1 — INFORMATIONAL log only, NOT a gate)
@@ -143,6 +160,7 @@ jobs:
          ref: ${{ github.event.repository.default_branch }}

      - name: Evaluate qa-review
+        id: eval
        env:
          GITEA_TOKEN: ${{ secrets.SOP_TIER_CHECK_TOKEN || secrets.GITHUB_TOKEN }}
          GITEA_HOST: git.moleculesai.app
@@ -157,3 +175,66 @@ jobs:
          REVIEW_CHECK_DEBUG: '0'
          REVIEW_CHECK_STRICT: '0'
        run: bash .gitea/scripts/review-check.sh
+
+      - name: Post required status context on pull_request_review
+        # Gitea Actions auto-publishes (pull_request_review) context
+        # for this event, but branch-protection requires (pull_request_target).
+        # We explicitly POST the BP-required context so the gate flips.
+        # Trust boundary: same BASE-ref script result, no PR-head code.
+        #
+        # TOKEN FIX (RC 8326): uses STATUS_POST_TOKEN (CTO-granted,
+        # msg d52cc72a). Dedicated narrow-scoped write:repository token
+        # for the explicit status POST. Evaluator step stays on
+        # SOP_TIER_CHECK_TOKEN (read-only) per deliberate security
+        # separation: eval computes, POST writes, never the same cred.
+        if: github.event_name == 'pull_request_review' && always()
+        env:
+          GITEA_TOKEN: ${{ secrets.STATUS_POST_TOKEN }}
+          GITEA_HOST: git.moleculesai.app
+          REPO: ${{ github.repository }}
+          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
+          EVAL_OUTCOME: ${{ steps.eval.outcome }}
+        run: |
+          set -euo pipefail
+          authfile=$(mktemp)
+          chmod 600 "$authfile"
+          printf 'header = "Authorization: token %s"\n' "$GITEA_TOKEN" > "$authfile"
+
+          prfile=$(mktemp)
+          code=$(curl -sS -o "$prfile" -w '%{http_code}' -K "$authfile" \
+            "https://${GITEA_HOST}/api/v1/repos/${REPO}/pulls/${PR_NUMBER}")
+          if [ "$code" != "200" ]; then
+            echo "::error::GET /pulls/${PR_NUMBER} returned HTTP ${code}"
+            rm -f "$prfile" "$authfile"
+            exit 1
+          fi
+          head_sha=$(jq -r '.head.sha // ""' "$prfile")
+          rm -f "$prfile"
+
+          if [ "$EVAL_OUTCOME" = "success" ]; then
+            status_state="success"
+            description="Approved via pull_request_review trigger"
+          else
+            status_state="failure"
+            description="Review check failed via pull_request_review trigger"
+          fi
+
+          body=$(jq -nc \
+            --arg state "$status_state" \
+            --arg context "qa-review / approved (pull_request_target)" \
+            --arg description "$description" \
+            '{state:$state, context:$context, description:$description}')
+
+          post_code=$(curl -sS -o /dev/null -w '%{http_code}' -X POST \
+            -K "$authfile" -H "Content-Type: application/json" \
+            -d "$body" \
+            "https://${GITEA_HOST}/api/v1/repos/${REPO}/statuses/${head_sha}")
+
+          rm -f "$authfile"
+
+          if [ "$post_code" != "200" ] && [ "$post_code" != "201" ]; then
+            echo "::error::POST /statuses/${head_sha} returned HTTP ${post_code}"
+            exit 1
+          fi
+
+          echo "::notice::posted ${status_state} for context=\"qa-review / approved (pull_request_target)\" on sha=${head_sha}"
@@ -6,25 +6,44 @@
 #
 # See `qa-review.yml` header for the full A1-α / A1.1 / A4 / A5 design
 # rationale; everything below is identical in shape.
+#
+# A1-α addendum (internal#760): review-event trigger added so the security
+# gate flips immediately when a team member submits an APPROVE review.
+# Uses `pull_request_review` types: [submitted] — verified live via
+# sop-tier-check.yml which provably fires this event (produces
+# `sop-tier-check / tier-check (pull_request_review)` contexts).
+# The job-level `if:` guard checks
+# `github.event.review.state == 'APPROVED' || 'approved'` so only APPROVE
+# reviews run the evaluator; COMMENT and REQUEST_CHANGES are skipped at
+# the job level. Branch-protection requires the `(pull_request_target)`
+# context variant, so the review-event path EXPLICITLY POSTS the required
+# context via the API. Trust boundary preserved (BASE ref, no PR-head).

 name: security-review

 on:
  pull_request_target:
    types: [opened, synchronize, reopened]
+  pull_request_review:
+    types: [submitted]

 permissions:
  contents: read
  pull-requests: read
-  secrets: read
+  statuses: write

 jobs:
  # bp-exempt: PR security review bot signal; required merge state is enforced by CI / all-required.
  approved:
-    # Comment-triggered refires live in review-refire-comments.yml. Keeping
-    # this workflow PR-only avoids comment-triggered queue storms.
+    # Gate the job:
+    #   - On pull_request_target events: always run.
+    #   - On pull_request_review_approved events: run so the gate flips
+    #     immediately when a team member submits an APPROVE review.
+    # Comment-triggered refires live in sop-checklist.yml review-refire job.
    if: |
-      github.event_name == 'pull_request_target'
+      github.event_name == 'pull_request_target' ||
+      (github.event_name == 'pull_request_review' &&
+       (github.event.review.state == 'APPROVED' || github.event.review.state == 'approved'))
    runs-on: ubuntu-latest
    steps:
      - name: Privilege check (A1.1 — INFORMATIONAL log only, NOT a gate)
@@ -57,6 +76,7 @@ jobs:
          ref: ${{ github.event.repository.default_branch }}

      - name: Evaluate security-review
+        id: eval
        env:
          GITEA_TOKEN: ${{ secrets.SOP_TIER_CHECK_TOKEN || secrets.GITHUB_TOKEN }}
          GITEA_HOST: git.moleculesai.app
@@ -68,3 +88,66 @@ jobs:
          REVIEW_CHECK_DEBUG: '0'
          REVIEW_CHECK_STRICT: '0'
        run: bash .gitea/scripts/review-check.sh
+
+      - name: Post required status context on pull_request_review
+        # Gitea Actions auto-publishes (pull_request_review) context
+        # for this event, but branch-protection requires (pull_request_target).
+        # We explicitly POST the BP-required context so the gate flips.
+        # Trust boundary: same BASE-ref script result, no PR-head code.
+        #
+        # TOKEN FIX (RC 8326): uses STATUS_POST_TOKEN (CTO-granted,
+        # msg d52cc72a). Dedicated narrow-scoped write:repository token
+        # for the explicit status POST. Evaluator step stays on
+        # SOP_TIER_CHECK_TOKEN (read-only) per deliberate security
+        # separation: eval computes, POST writes, never the same cred.
+        if: github.event_name == 'pull_request_review' && always()
+        env:
+          GITEA_TOKEN: ${{ secrets.STATUS_POST_TOKEN }}
+          GITEA_HOST: git.moleculesai.app
+          REPO: ${{ github.repository }}
+          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
+          EVAL_OUTCOME: ${{ steps.eval.outcome }}
+        run: |
+          set -euo pipefail
+          authfile=$(mktemp)
+          chmod 600 "$authfile"
+          printf 'header = "Authorization: token %s"\n' "$GITEA_TOKEN" > "$authfile"
+
+          prfile=$(mktemp)
+          code=$(curl -sS -o "$prfile" -w '%{http_code}' -K "$authfile" \
+            "https://${GITEA_HOST}/api/v1/repos/${REPO}/pulls/${PR_NUMBER}")
+          if [ "$code" != "200" ]; then
+            echo "::error::GET /pulls/${PR_NUMBER} returned HTTP ${code}"
+            rm -f "$prfile" "$authfile"
+            exit 1
+          fi
+          head_sha=$(jq -r '.head.sha // ""' "$prfile")
+          rm -f "$prfile"
+
+          if [ "$EVAL_OUTCOME" = "success" ]; then
+            status_state="success"
+            description="Approved via pull_request_review trigger"
+          else
+            status_state="failure"
+            description="Review check failed via pull_request_review trigger"
+          fi
+
+          body=$(jq -nc \
+            --arg state "$status_state" \
+            --arg context "security-review / approved (pull_request_target)" \
+            --arg description "$description" \
+            '{state:$state, context:$context, description:$description}')
+
+          post_code=$(curl -sS -o /dev/null -w '%{http_code}' -X POST \
+            -K "$authfile" -H "Content-Type: application/json" \
+            -d "$body" \
+            "https://${GITEA_HOST}/api/v1/repos/${REPO}/statuses/${head_sha}")
+
+          rm -f "$authfile"
+
+          if [ "$post_code" != "200" ] && [ "$post_code" != "201" ]; then
+            echo "::error::POST /statuses/${head_sha} returned HTTP ${post_code}"
+            exit 1
+          fi
+
+          echo "::notice::posted ${status_state} for context=\"security-review / approved (pull_request_target)\" on sha=${head_sha}"
@@ -179,10 +179,10 @@ jobs:
      - name: Refire qa-review status
        if: steps.classify.outputs.run_qa == 'true'
        env:
-          # RFC_324_TEAM_READ_TOKEN is read-only (team membership read scope only).
-          # review-refire-status.sh POSTs to /statuses — requires write scope.
-          # SOP_TIER_CHECK_TOKEN carries write:repository + write:issue + read:organization.
+          # Evaluator (review-check.sh + GET /pulls) stays on read-scoped token.
          GITEA_TOKEN: ${{ secrets.SOP_TIER_CHECK_TOKEN || secrets.GITHUB_TOKEN }}
+          # Explicit POST /statuses uses narrow-scoped write:repository token.
+          STATUS_POST_TOKEN: ${{ secrets.STATUS_POST_TOKEN }}
          GITEA_HOST: git.moleculesai.app
          REPO: ${{ github.repository }}
          PR_NUMBER: ${{ github.event.issue.number }}
@@ -198,10 +198,10 @@ jobs:
      - name: Refire security-review status
        if: steps.classify.outputs.run_security == 'true'
        env:
-          # RFC_324_TEAM_READ_TOKEN is read-only (team membership read scope only).
-          # review-refire-status.sh POSTs to /statuses — requires write scope.
-          # SOP_TIER_CHECK_TOKEN carries write:repository + write:issue + read:organization.
+          # Evaluator (review-check.sh + GET /pulls) stays on read-scoped token.
          GITEA_TOKEN: ${{ secrets.SOP_TIER_CHECK_TOKEN || secrets.GITHUB_TOKEN }}
+          # Explicit POST /statuses uses narrow-scoped write:repository token.
+          STATUS_POST_TOKEN: ${{ secrets.STATUS_POST_TOKEN }}
          GITEA_HOST: git.moleculesai.app
          REPO: ${{ github.repository }}
          PR_NUMBER: ${{ github.event.issue.number }}
@@ -60,6 +60,7 @@ concurrency:
  cancel-in-progress: true

 jobs:
+  # bp-required: pending #718 — soak-then-promote, not in BP yet.
  compare:
    name: Compare synced providers.yaml against controlplane canonical
    runs-on: ubuntu-latest
@@ -67,6 +67,7 @@ concurrency:
  cancel-in-progress: true

 jobs:
+  # bp-required: pending #718 — soak-then-promote, not in BP yet.
  verify:
    name: Regenerate providers artifact and fail on drift
    runs-on: ubuntu-latest
@@ -60,11 +60,26 @@ test.describe("MobileChat", () => {

    await expect(page.getByText("Echo: Mobile persistence")).toBeVisible({ timeout: 15_000 });

+    // Reload and deterministically wait for the chat-history GET that
+    // rehydrates the transcript to come back 2xx, rather than racing a
+    // fixed-timeout render assertion against an in-flight fetch. The
+    // server now persists the a2a_receive row SYNCHRONOUSLY before the
+    // send's 200 (workspace-server logA2ASuccess), so the row is
+    // guaranteed present by the time this GET runs — the wait is for
+    // hydration latency, not for a still-racing write.
+    const historyResponse = page.waitForResponse(
+      (resp) =>
+        resp.url().includes("/chat-history") &&
+        resp.request().method() === "GET" &&
+        resp.status() === 200,
+      { timeout: 15_000 },
+    );
    await page.reload();
    await page.waitForSelector("[data-testid='chat-panel']", { timeout: 10_000 });
+    await historyResponse;

-    await expect(page.getByText("Mobile persistence", { exact: true })).toBeVisible({ timeout: 5_000 });
-    await expect(page.getByText("Echo: Mobile persistence")).toBeVisible({ timeout: 5_000 });
+    await expect(page.getByText("Mobile persistence", { exact: true })).toBeVisible();
+    await expect(page.getByText("Echo: Mobile persistence")).toBeVisible();
  });

  test("composer auto-grows with multi-line text", async ({ page }) => {
@@ -250,7 +250,38 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
  const workspaceId = ws.body.id as string;
  console.log(`[staging-setup] Workspace created: ${workspaceId}`);

-  // 6. Wait for workspace online
+  // 6. Wait for workspace RENDERABLE.
+  //
+  // This harness exists to verify the canvas *tab UI* renders (staging-
+  // tabs.spec.ts: open each of the 13 workspace-panel tabs, assert no hard
+  // crash / no "Failed to load" toast). It does NOT exercise the agent —
+  // no LLM call is made, the spec even mocks /cp/auth/me and 401→200. All
+  // it needs is a workspace ROW that the canvas lists so the node renders
+  // and the side-panel tabs open. A fully-`online` agent is NOT required.
+  //
+  // That distinction became load-bearing on 2026-06-03: workspace-server
+  // #2162 (fix(provision): platform-managed workspace must fail-closed when
+  // CP proxy env absent) made a platform_managed workspace ABORT AT BOOT
+  // with MISSING_PLATFORM_PROXY when MOLECULE_LLM_BASE_URL /
+  // MOLECULE_LLM_USAGE_TOKEN are not present in the tenant's env. The
+  // canvas E2E creates a bare hermes/gpt-4o workspace, which defaults
+  // closed to platform_managed (workspace_provision.go:~1009), and the
+  // staging tenant does not carry the CP proxy env — so the agent never
+  // starts. Pre-#2162 this same workspace booted credential-less (the bug
+  // #2162 fixed) and the tabs rendered fine; #2162 is a correct production
+  // safety fix, but it surfaced here as `status:"failed", uptime_seconds:0,
+  // last_sample_error:null` — the pre-start credential-abort shape — and the
+  // old hard-throw turned a UI-irrelevant boot skip into a main-red
+  // (core#2199). The agent boot stage is simply not what this test gates.
+  //
+  // So: online is the happy path. A `failed` row that is the PRE-START
+  // credential-abort shape (the agent process never ran: uptime_seconds==0
+  // AND no last_sample_error) is treated as RENDERABLE — the row exists,
+  // the node + tabs render, proceed. We do NOT mask a real boot regression:
+  // any `failed` carrying a last_sample_error, OR a non-zero uptime (the
+  // agent started then crashed — image pull, panic, PYTHONPATH, etc.),
+  // still hard-throws. Genuine *infra* provision failure is already caught
+  // loud one step earlier at the org level (instance_status === "failed").
  await waitFor<boolean>(
    async () => {
      const r = await jsonFetch(`${tenantURL}/workspaces/${workspaceId}`, {
@@ -259,6 +290,24 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
      if (r.status !== 200) return null;
      if (r.body?.status === "online") return true;
      if (r.body?.status === "failed") {
+        const uptime = Number(r.body?.uptime_seconds ?? 0);
+        const sampleErr = r.body?.last_sample_error;
+        const preStartCredentialAbort = uptime === 0 && !sampleErr;
+        if (preStartCredentialAbort) {
+          // Agent never started (no LLM cred on this staging tenant — the
+          // expected #2162 platform-proxy gap). The workspace row still
+          // renders, which is all the tab-UI test needs. Proceed, but log
+          // loudly so a real "agent never booted because of something else"
+          // is not silently normalized.
+          console.warn(
+            `[staging-setup] workspace ${workspaceId} is 'failed' with the pre-start ` +
+              `credential-abort shape (uptime_seconds=0, no last_sample_error) — agent did ` +
+              `not boot (expected on staging without CP LLM proxy env, post workspace-server ` +
+              `#2162). The tab-UI test does not exercise the agent; proceeding with the ` +
+              `workspace row, which renders regardless. full body: ${JSON.stringify(r.body)}`,
+          );
+          return true;
+        }
        // last_sample_error is often empty when the failure happens before
        // the agent emits a sample (e.g. boot crash, image pull error,
        // missing PYTHONPATH, OpenAI quota at startup). Dumping the full
@@ -266,8 +315,8 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
        // needs without a second probe. Otherwise this propagates as a
        // bare "Workspace failed: " — the exact useless message that
        // sent #2632 to the issue tracker.
-        const detail = r.body.last_sample_error
-          ? r.body.last_sample_error
+        const detail = sampleErr
+          ? sampleErr
          : `(no last_sample_error) full body: ${JSON.stringify(r.body)}`;
        throw new Error(`Workspace failed: ${detail}`);
      }
@@ -277,7 +326,7 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
    10_000,
    "workspace online",
  );
-  console.log(`[staging-setup] Workspace online`);
+  console.log(`[staging-setup] Workspace renderable`);

  // 7. Hand state off to tests + teardown — overwrite the slug-only
  // bootstrap state with the full state spec tests need.
@@ -8,9 +8,12 @@ import { ExternalConnectModal, type ExternalConnectionInfo } from "./ExternalCon
 import {
  ProviderModelSelector,
  buildProviderCatalog,
+  buildProviderCatalogFromRegistry,
  findProviderForModel,
  type SelectorModel,
  type SelectorValue,
+  type RegistryProvider,
+  type RegistryModel,
 } from "./ProviderModelSelector";

 interface WorkspaceOption {
@@ -32,6 +35,16 @@ interface TemplateSpec {
  model?: string;
  models?: SelectorModel[];
  providers?: string[];
+  // internal#718 P3 registry-served fields (additive; absent on older
+  // backends and for non-registry runtimes). When registry_backed is true the
+  // provider→model catalog is built from registry_providers/registry_models so
+  // each model's DERIVED provider (e.g. moonshot/kimi-k2.6 → "platform") drives
+  // the dropdown bucket and the create payload's llm_provider — instead of the
+  // legacy inferVendor heuristic that slash-splits the id into "moonshot".
+  // Mirrors ConfigTab's RuntimeOption loader (RFC#340 Fix C).
+  registry_backed?: boolean;
+  registry_providers?: RegistryProvider[];
+  registry_models?: RegistryModel[];
 }

 const DEFAULT_RUNTIME = "claude-code";
@@ -168,15 +181,53 @@ export function CreateWorkspaceButton() {
    }),
    [runtime, templateSpecs],
  );
-  const llmModels = useMemo(
-    () => {
-      const sourceSpec = selectedTemplateSpec ?? selectedRuntimeTemplateSpec;
-      if (!sourceSpec?.models?.length) return [];
-      return sourceSpec.models;
-    },
+  // The /templates row backing the LLM picker: an explicitly-selected
+  // workspace template wins, else the base runtime template row.
+  const llmSourceSpec = useMemo<TemplateSpec | null>(
+    () => selectedTemplateSpec ?? selectedRuntimeTemplateSpec,
    [selectedRuntimeTemplateSpec, selectedTemplateSpec],
  );
-  const llmCatalog = useMemo(() => buildProviderCatalog(llmModels), [llmModels]);
+  // internal#718 P3 / RFC#340 Fix C: a runtime is registry-backed when the
+  // /templates row says so AND it served a non-empty registry_models set.
+  // Mirrors ConfigTab's `registryBacked` derivation exactly.
+  const registryBacked = useMemo(
+    () =>
+      llmSourceSpec?.registry_backed === true &&
+      (llmSourceSpec.registry_models?.length ?? 0) > 0,
+    [llmSourceSpec],
+  );
+  // Models fed to the selector dropdown. For a registry-backed runtime use the
+  // registry-served native set, carrying each model's DERIVED provider so the
+  // selector buckets it correctly (moonshot/kimi-k2.6 → "platform", not the
+  // inferVendor "moonshot"). Otherwise fall back to the template-served
+  // models[] + the legacy heuristic — same fallback ConfigTab keeps.
+  const llmModels = useMemo<SelectorModel[]>(
+    () => {
+      if (registryBacked) {
+        return (llmSourceSpec?.registry_models ?? []).map((m) => ({
+          id: m.id,
+          name: m.name,
+          ...(m.provider ? { provider: m.provider } : {}),
+        }));
+      }
+      return llmSourceSpec?.models?.length ? llmSourceSpec.models : [];
+    },
+    [registryBacked, llmSourceSpec],
+  );
+  // Registry-backed path: build the catalog from registry_providers/
+  // registry_models so dropdown labels + billing + the derived provider come
+  // from the provider-registry SSOT (restores the "Platform" bucket). Legacy
+  // path: re-infer from models[] via buildProviderCatalog (inferVendor).
+  const llmCatalog = useMemo(
+    () =>
+      registryBacked
+        ? buildProviderCatalogFromRegistry(
+            llmSourceSpec?.registry_providers ?? [],
+            llmSourceSpec?.registry_models ?? [],
+          )
+        : buildProviderCatalog(llmModels),
+    [registryBacked, llmSourceSpec, llmModels],
+  );
  const selectedLLMProvider = useMemo(
    () => llmCatalog.find((p) => p.id === llmSelection.providerId) ?? llmCatalog[0],
    [llmCatalog, llmSelection.providerId],
@@ -184,7 +235,7 @@ export function CreateWorkspaceButton() {

  useEffect(() => {
    if (llmCatalog.length === 0) return;
-    const sourceDefault = (selectedTemplateSpec ?? selectedRuntimeTemplateSpec)?.model?.trim();
+    const sourceDefault = llmSourceSpec?.model?.trim();
    const platformProvider = llmCatalog.find((p) => p.vendor === "platform");
    const matched = sourceDefault ? findProviderForModel(llmCatalog, sourceDefault) : null;
    const next = platformProvider ?? matched ?? llmCatalog[0];
@@ -197,7 +248,7 @@ export function CreateWorkspaceButton() {
      envVars: next.envVars,
    });
    setLLMSecret("");
-  }, [llmCatalog, selectedRuntimeTemplateSpec, selectedTemplateSpec]);
+  }, [llmCatalog, llmSourceSpec]);

  // Reset form and load workspaces whenever dialog opens
  useEffect(() => {
@@ -461,6 +512,7 @@ export function CreateWorkspaceButton() {
                </div>
                <ProviderModelSelector
                  models={llmModels}
+                  catalog={registryBacked ? llmCatalog : undefined}
                  value={llmSelection}
                  onChange={(next) => {
                    setLLMSelection(next);
@@ -454,6 +454,128 @@ describe("CreateWorkspaceDialog — dynamic runtime provider picker", () => {
  });
 });

+// ---------------------------------------------------------------------------
+// Registry-backed provider catalog (RFC#340 Fix C)
+//
+// Regression guard for the mis-bucketing bug: when a registry-backed
+// claude-code template serves `moonshot/kimi-k2.6` whose DERIVED provider is
+// `platform`, the dialog must build the dropdown from registry_providers/
+// registry_models (buildProviderCatalogFromRegistry) — NOT the legacy
+// inferVendor heuristic which slash-splits the id into "moonshot". The
+// distinguishing trait of this fixture: the plain `models[]` array does NOT
+// carry an explicit `provider` field, so the LEGACY path would bucket the
+// model under "moonshot" and send llm_provider:"moonshot". Only the
+// registry-backed path yields the Platform bucket + llm_provider:"platform".
+// ---------------------------------------------------------------------------
+
+// claude-code template whose plain models[] is UN-annotated (no explicit
+// provider). The derived-provider annotation lives ONLY in registry_models.
+const REGISTRY_TEMPLATE = {
+  id: "claude-code-default",
+  name: "Claude Code Agent",
+  runtime: "claude-code",
+  model: "moonshot/kimi-k2.6",
+  // Legacy fields — note: NO explicit provider on the platform model, so the
+  // legacy inferVendor path would slash-split it into "moonshot".
+  providers: ["platform", "minimax", "anthropic"],
+  models: [
+    { id: "moonshot/kimi-k2.6", name: "Kimi K2.6", required_env: [] },
+    { id: "MiniMax-M2.7", name: "MiniMax M2.7", required_env: ["MINIMAX_API_KEY"] },
+    { id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6", required_env: ["ANTHROPIC_API_KEY"] },
+  ],
+  // Registry-served SSOT (internal#718 P3). DeriveProvider resolved
+  // moonshot/kimi-k2.6 → "platform"; MiniMax-M2.7 → "minimax".
+  registry_backed: true,
+  registry_providers: [
+    { name: "platform", display_name: "Platform", auth_env: [], billing_mode: "platform_managed" },
+    { name: "minimax", display_name: "MiniMax", auth_env: ["MINIMAX_API_KEY"], billing_mode: "byok" },
+    { name: "anthropic", display_name: "Anthropic API", auth_env: ["ANTHROPIC_API_KEY"], billing_mode: "byok" },
+  ],
+  registry_models: [
+    { id: "moonshot/kimi-k2.6", name: "Kimi K2.6", provider: "platform", billing_mode: "platform_managed" },
+    { id: "MiniMax-M2.7", name: "MiniMax M2.7", provider: "minimax", billing_mode: "byok" },
+    { id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6", provider: "anthropic", billing_mode: "byok" },
+  ],
+};
+
+describe("CreateWorkspaceDialog — registry-backed provider catalog (RFC#340 Fix C)", () => {
+  beforeEach(() => {
+    mockGet.mockImplementation(async (url: string) => {
+      if (url === "/templates") {
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        return [REGISTRY_TEMPLATE] as any;
+      }
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      return SAMPLE_WORKSPACES as any;
+    });
+  });
+
+  it("shows the Platform provider bucket for the registry-backed claude-code runtime", async () => {
+    await openDialog();
+    const providerSelect = await waitFor(() => {
+      const sel = document.querySelector("[data-testid='provider-select']") as HTMLSelectElement;
+      expect(sel).toBeTruthy();
+      return sel;
+    });
+    const labels = Array.from(providerSelect.options).map((o) => o.text.trim());
+    // Registry display_name "Platform" appears — NOT "moonshot" from the
+    // legacy slash-split heuristic.
+    expect(labels).toContain("Platform");
+    expect(labels).not.toContain("moonshot");
+    // Bucket id is the registry-keyed id, vendor is the bare provider name.
+    const values = Array.from(providerSelect.options).map((o) => o.value);
+    expect(values).toContain("registry|platform");
+  });
+
+  it("sends llm_provider: platform (not moonshot) for moonshot/kimi-k2.6", async () => {
+    await openDialog();
+    fireEvent.change(screen.getByPlaceholderText("e.g. SEO Agent"), {
+      target: { value: "Kimi Agent" },
+    });
+    // Wait for the registry default to settle on the Platform bucket + model.
+    await waitFor(() => {
+      const modelSelect = document.querySelector("[data-testid='model-select']") as HTMLSelectElement;
+      expect(modelSelect?.value).toBe("moonshot/kimi-k2.6");
+    });
+
+    const createBtn = screen.getAllByRole("button").find((b) => b.textContent === "Create");
+    fireEvent.click(createBtn!);
+
+    await waitFor(() => expect(mockPost).toHaveBeenCalled());
+    const body = mockPost.mock.calls[0][1] as Record<string, unknown>;
+    expect(body.model).toBe("moonshot/kimi-k2.6");
+    expect(body.llm_provider).toBe("platform");
+    // Platform is auth-env-free → no BYOK secret.
+    expect(body.secrets).toBeUndefined();
+  });
+
+  it("buckets MiniMax-M2.7 under its derived provider and sends llm_provider: minimax", async () => {
+    await openDialog();
+    fireEvent.change(screen.getByPlaceholderText("e.g. SEO Agent"), {
+      target: { value: "MiniMax Agent" },
+    });
+    await waitFor(() => {
+      const sel = document.querySelector("[data-testid='provider-select']") as HTMLSelectElement;
+      expect(Array.from(sel.options).map((o) => o.value)).toContain("registry|minimax");
+    });
+    fireEvent.change(document.querySelector("[data-testid='provider-select']") as HTMLSelectElement, {
+      target: { value: "registry|minimax" },
+    });
+    fireEvent.change(document.getElementById("llm-secret-input") as HTMLInputElement, {
+      target: { value: "sk-minimax-test" },
+    });
+
+    const createBtn = screen.getAllByRole("button").find((b) => b.textContent === "Create");
+    fireEvent.click(createBtn!);
+
+    await waitFor(() => expect(mockPost).toHaveBeenCalled());
+    const body = mockPost.mock.calls[0][1] as Record<string, unknown>;
+    expect(body.model).toBe("MiniMax-M2.7");
+    expect(body.llm_provider).toBe("minimax");
+    expect(body.secrets).toEqual({ MINIMAX_API_KEY: "sk-minimax-test" });
+  });
+});
+
 // ---------------------------------------------------------------------------
 // budget_limit field tests (#541)
 // ---------------------------------------------------------------------------
@@ -34,8 +34,6 @@
  ],
  "org_templates": [
    {"name": "molecule-dev", "repo": "molecule-ai/molecule-ai-org-template-molecule-dev", "ref": "main"},
-    {"name": "free-beats-all", "repo": "molecule-ai/molecule-ai-org-template-free-beats-all", "ref": "main"},
-    {"name": "medo-smoke", "repo": "molecule-ai/molecule-ai-org-template-medo-smoke", "ref": "main"},
    {"name": "molecule-worker-gemini", "repo": "molecule-ai/molecule-ai-org-template-molecule-worker-gemini", "ref": "main"},
    {"name": "ux-ab-lab", "repo": "molecule-ai/molecule-ai-org-template-ux-ab-lab", "ref": "main"}
  ]
@@ -0,0 +1,131 @@
+# Developer SOP — PR review gate auto-fire and stale-head handling
+
+> Last updated: 2026-06-03 (cp#2159 follow-up)
+>
+> Applies to: all core-PR authors and reviewers on `molecule-core` and sibling
+> repos using the `qa-review` + `security-review` branch-protection gates.
+
+---
+
+## 1. Gitea PR-head workflow-selection rule
+
+**Rule:** For `pull_request_target` and `pull_request_review` events, Gitea
+loads the workflow definition from the **PR's HEAD branch**, not from the
+base (`main`) branch.
+
+This is different from GitHub Actions, where `pull_request_target` always
+loads workflows from the base branch. Gitea's behaviour means:
+
+- A PR that was opened **before** the `pull_request_review` trigger was added
+to `qa-review.yml` / `security-review.yml` will **NOT** auto-fire on review,
+because its HEAD still contains the old workflow YAML (no trigger).
+
+- A PR that was opened **after** the trigger was added (or that has been
+rebased onto a commit containing the trigger) **WILL** auto-fire, because its
+HEAD contains the new workflow YAML.
+
+### Ops implication
+
+| PR head contains `pull_request_review` trigger? | Behaviour on APPROVED review |
+|---|---|
+| **Yes** (cut from current main, or rebased) | Workflows auto-queue, evaluate, and POST the `(pull_request_target)` context automatically. No slash-command needed. |
+| **No** (stale head, opened before #2157) | Nothing fires. Use `/qa-recheck` + `/security-recheck` slash-commands in a PR comment, OR rebase onto current main. |
+
+---
+
+## 2. Standard core-PR flow (post-#2157)
+
+```
+1. Author opens PR from a branch based on current main
+   → qa-review + security-review workflows run on pull_request_target
+   → status contexts post (initial eval, usually red until reviews land)
+
+2. Reviewers submit real APPROVED reviews
+   → If PR head has the trigger: workflows AUTO-FIRE on pull_request_review
+   → Contexts flip green (or stay red if reviewer is not in team)
+
+3. [Optional] If contexts did not flip (stale head, event lost, etc.):
+   → Anyone can comment `/qa-recheck` or `/security-recheck`
+   → sop-checklist.yml refires the evaluator (read-only, idempotent)
+
+4. Both qa-review + security-review contexts are green
+   → Plain Do:merge (no force-merge needed)
+```
+
+### Key point
+
+The `/qa-recheck` and `/security-recheck` commands are a **backstop**, not the
+primary path. PRs cut from current main should auto-fire without manual
+intervention.
+
+---
+
+## 3. Diagnosing a stale head
+
+If a PR has real team-member APPROVED reviews but the qa/security contexts
+remain red and no workflow run appears on the PR's "Actions" tab for the
+review event, the PR head is likely stale.
+
+### Quick check
+
+```bash
+# From the PR page, look at the head commit SHA, then:
+curl -sS "https://git.moleculesai.app/api/v1/repos/molecule-ai/molecule-core/contents/.gitea/workflows/qa-review.yml?ref=<HEAD_SHA>" \
+  | jq -r '.content' | base64 -d | grep -c 'pull_request_review'
+# 0  → stale head (no trigger in that version of the workflow)
+# >0 → trigger present; auto-fire SHOULD work (if it didn't, file a tracker)
+```
+
+### Automated diagnostic
+
+The test suite includes `test_gate_stale_head_diagnostic.py`, which reports
+"auto-fire impossible for this PR" when the head lacks the trigger. Run it
+in CI or locally with:
+
+```bash
+PR_NUMBER=123 python -m pytest .gitea/scripts/tests/test_gate_stale_head_diagnostic.py -v
+```
+
+---
+
+## 4. Rebasing vs. slash-refire
+
+| Approach | When to use | Trade-off |
+|---|---|---|
+| **Rebase onto current main** | PR is genuinely stale (head lacks trigger OR head is far behind main) | Clean history, gets all recent fixes, but requires force-push and re-approval if the branch was protected |
+| **`/qa-recheck` + `/security-recheck`** | PR head is recent but the review event was missed, or you want to avoid rebase churn | Quick, no force-push, but does NOT fix a missing trigger in the head |
+
+**Do not** use slash-refire as a substitute for rebasing a stale head. If the
+workflow YAML in the PR head does not contain `pull_request_review`, no amount
+of rechecking will make auto-fire work.
+
+---
+
+## 5. Live-fire verification
+
+The `test_gate_auto_fire_live.py` regression test exercises the full runtime
+path: it submits an APPROVED review to a test PR and polls for the
+`(pull_request_target)` status contexts. It is skipped when no API token is
+available, and is intended to catch runtime non-fire that static structural
+tests (e.g. `test_gate_review_auto_fire.py`) cannot detect.
+
+Run manually with:
+
+```bash
+export GITEA_HOST=git.moleculesai.app
+export GITEA_TOKEN=<your-token>
+export REPO=molecule-ai/molecule-core
+export LIVEFIRE_PR_NUMBER=<test-pr-number>
+python -m pytest .gitea/scripts/tests/test_gate_auto_fire_live.py -v
+```
+
+---
+
+## References
+
+- #2159 — gate auto-trigger not firing (root cause: stale PR heads lacking
+the `pull_request_review` trigger, NOT a workflow code defect)
+- #765 — static structural regression test for gate configuration
+- #2157 — merged trigger addition (`pull_request_review` types: [submitted])
+- #2020 — milestone confirming gate infrastructure is stable
+- RFC#324 — qa-review + security-review design
@@ -23,23 +23,61 @@
 #                  their provider entries, otherwise the workspace boots
 #                  reachable but the first A2A call hits the wrong auth path.
 #
-# When E2E_MODEL_SLUG is set, it overrides this dispatch — useful when an
-# operator dispatches the workflow to test a specific slug.
+# PLATFORM-MANAGED path (E2E_LLM_PATH=platform) — the moonshot/kimi
+# NOT_CONFIGURED regression (RFC#340 Fix A #2187):
+#
+#   The branches above all exercise BYOK: a tenant key (MINIMAX/ANTHROPIC/
+#   OPENAI) is injected as a workspace secret and the model id resolves to that
+#   vendor's *BYOK* provider entry. That path NEVER exercises the platform arm —
+#   the exact arm that booted "moonshot/kimi-k2.6" into NOT_CONFIGURED in prod,
+#   because the generated config.yaml lacked the derived `provider: platform`.
+#
+#   E2E_LLM_PATH=platform selects a platform-managed model id (slash-namespaced,
+#   no tenant key — Molecule owns billing via the CP LLM proxy). The default is
+#   "moonshot/kimi-k2.6", the headline incident combo. Override the specific
+#   platform model with E2E_MODEL_SLUG. The provision branch in
+#   test_staging_full_saas.sh sends NO secrets for this path (platform-managed
+#   needs none), so the workspace must boot online purely on the proxy env the
+#   control plane injects + the manifest-derived `provider: platform` that Fix A
+#   stamps. That is the REAL boot-path assertion the deterministic unit test
+#   (workspace_provision_platform_boot_test.go) cannot make.
+#
+# When E2E_MODEL_SLUG is set, it overrides this dispatch entirely — useful when
+# an operator dispatches the workflow to test a specific slug (or a specific
+# platform model id).
 #
 # Unit tested by tests/e2e/test_model_slug.sh — every branch must stay
 # pinned because regressions silently mask as "Could not resolve
 # authentication method" + the synth-E2E gate goes red without naming
 # the slug-format mismatch.

+# Default platform-managed model for the platform-boot regression path. The
+# exact id that booted NOT_CONFIGURED in prod. Must stay a member of the
+# claude-code `platform` arm in workspace-server/internal/providers/providers.yaml
+# (the deterministic suite TestEnsureDefaultConfig_StampsProviderForEverySSOTPlatformModel
+# enforces every member of that arm derives provider=platform). Resolved INSIDE
+# pick_model_slug via ${E2E_DEFAULT_PLATFORM_MODEL:-...} so callers can override
+# it (or unset it) without tripping `set -u`.
+E2E_DEFAULT_PLATFORM_MODEL_FALLBACK="moonshot/kimi-k2.6"
+
 # Usage: pick_model_slug <runtime>
 #   stdout: the slug string
 #   E2E_MODEL_SLUG (env): if set + non-empty, used as-is (operator override)
+#   E2E_LLM_PATH=platform (env): select the platform-managed model id
+#     (E2E_DEFAULT_PLATFORM_MODEL) instead of a BYOK slug. Takes precedence over
+#     the per-key BYOK branches; E2E_MODEL_SLUG still wins over everything.
 pick_model_slug() {
  local runtime="${1:-}"
  if [ -n "${E2E_MODEL_SLUG:-}" ]; then
    printf '%s' "$E2E_MODEL_SLUG"
    return 0
  fi
+  # Platform-managed path: the slash-namespaced platform model, no tenant key.
+  # Exercises the arm the moonshot/kimi NOT_CONFIGURED bug shipped on.
+  if [ "${E2E_LLM_PATH:-}" = "platform" ]; then
+    printf '%s' "${E2E_DEFAULT_PLATFORM_MODEL:-$E2E_DEFAULT_PLATFORM_MODEL_FALLBACK}"
+    return 0
+  fi
  case "$runtime" in
    hermes)      printf 'openai/gpt-4o' ;;
    claude-code)
@@ -65,6 +65,28 @@ assert_eq "claude-code + both keys → MiniMax priority"            "$got" "Mini
 run_test "unknown runtime → slash-form fallback"                   gemini      "openai/gpt-4o"
 run_test "empty runtime → slash-form fallback"                     ""          "openai/gpt-4o"

+# ── Platform-managed path (E2E_LLM_PATH=platform) ──
+# The moonshot/kimi NOT_CONFIGURED regression path (RFC#340 Fix A #2187).
+# Selects the slash-namespaced platform model (default moonshot/kimi-k2.6),
+# takes precedence over the per-key BYOK branches, and is itself overridden by
+# E2E_MODEL_SLUG. These pins guard the harness's ability to drive the platform
+# arm — the one the prod bug shipped on.
+echo
+echo "Test: pick_model_slug — platform-managed path (E2E_LLM_PATH=platform)"
+echo
+
+got=$(unset E2E_MODEL_SLUG E2E_DEFAULT_PLATFORM_MODEL; E2E_LLM_PATH=platform pick_model_slug claude-code)
+assert_eq "claude-code + platform path → headline kimi model"      "$got" "moonshot/kimi-k2.6"
+
+got=$(unset E2E_MODEL_SLUG E2E_DEFAULT_PLATFORM_MODEL; E2E_LLM_PATH=platform E2E_MINIMAX_API_KEY="mx-stray" pick_model_slug claude-code)
+assert_eq "platform path beats a stray BYOK key (no mask)"         "$got" "moonshot/kimi-k2.6"
+
+got=$(unset E2E_MODEL_SLUG; E2E_LLM_PATH=platform E2E_DEFAULT_PLATFORM_MODEL="minimax/MiniMax-M3" pick_model_slug claude-code)
+assert_eq "platform path honours E2E_DEFAULT_PLATFORM_MODEL"        "$got" "minimax/MiniMax-M3"
+
+got=$(unset E2E_DEFAULT_PLATFORM_MODEL; E2E_MODEL_SLUG="anthropic/claude-opus-4-7" E2E_LLM_PATH=platform pick_model_slug claude-code)
+assert_eq "E2E_MODEL_SLUG still wins over platform path"            "$got" "anthropic/claude-opus-4-7"
+
 # ── Override via E2E_MODEL_SLUG ──
 # When the operator sets E2E_MODEL_SLUG, the per-runtime dispatch is
 # bypassed. Used during workflow_dispatch to A/B specific slugs.
@@ -234,9 +234,30 @@ elif [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
  SECRETS_JSON=$(python3 -c "import json,os;k=os.environ['E2E_OPENAI_API_KEY'];print(json.dumps({'OPENAI_API_KEY':k,'OPENAI_BASE_URL':'https://api.openai.com/v1','MODEL_PROVIDER':'openai:gpt-4o','HERMES_INFERENCE_PROVIDER':'custom','HERMES_CUSTOM_BASE_URL':'https://api.openai.com/v1','HERMES_CUSTOM_API_KEY':k,'HERMES_CUSTOM_API_MODE':'chat_completions'}))")
 fi

+# Workspace-create now enforces the MODEL_REQUIRED contract: there is NO
+# platform-side default model for a runtime (feedback_workspace_model_required_
+# no_platform_default). Every create MUST carry an explicit `model`, or the CP
+# rejects it with MODEL_REQUIRED before this gate's peer-visibility assertion
+# can run. We pick a PLATFORM-MANAGED id (Molecule owns billing — no tenant key
+# needed; this gate only needs the workspace to boot + list peers, not heavy
+# LLM work), validated against the controlplane providers SSOT
+# (internal/providers/providers.yaml runtimes.<rt>.providers[platform].models):
+#   claude-code → anthropic/claude-sonnet-4-6   (platform claude model)
+#   hermes/openclaw → moonshot/kimi-k2.6         (their only platform family)
+# E2E_MODEL_SLUG overrides for operator-dispatched runs.
+pv_platform_model_for_runtime() {
+  if [ -n "${E2E_MODEL_SLUG:-}" ]; then printf '%s' "$E2E_MODEL_SLUG"; return 0; fi
+  case "$1" in
+    claude-code) printf 'anthropic/claude-sonnet-4-6' ;;
+    hermes|openclaw) printf 'moonshot/kimi-k2.6' ;;
+    *) printf 'moonshot/kimi-k2.6' ;;
+  esac
+}
+
 log "4/6 provisioning parent (claude-code) + one sibling per runtime under test..."
+PARENT_MODEL=$(pv_platform_model_for_runtime claude-code)
 P_RESP=$(tenant_call POST /workspaces \
-  -d "{\"name\":\"pv-parent\",\"runtime\":\"claude-code\",\"tier\":3,\"secrets\":$SECRETS_JSON}")
+  -d "{\"name\":\"pv-parent\",\"runtime\":\"claude-code\",\"model\":\"$PARENT_MODEL\",\"tier\":3,\"secrets\":$SECRETS_JSON}")
 PARENT_ID=$(echo "$P_RESP" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null)
 [ -n "$PARENT_ID" ] || fail "parent create failed: $(echo "$P_RESP" | head -c 300)"
 log "    PARENT_ID=$PARENT_ID"
@@ -245,8 +266,9 @@ log "    PARENT_ID=$PARENT_ID"
 declare -A WS_IDS WS_TOKENS
 ALL_WS_IDS="$PARENT_ID"
 for rt in $PV_RUNTIMES; do
+  RT_MODEL=$(pv_platform_model_for_runtime "$rt")
  R=$(tenant_call POST /workspaces \
-    -d "{\"name\":\"pv-$rt\",\"runtime\":\"$rt\",\"tier\":2,\"parent_id\":\"$PARENT_ID\",\"secrets\":$SECRETS_JSON}")
+    -d "{\"name\":\"pv-$rt\",\"runtime\":\"$rt\",\"model\":\"$RT_MODEL\",\"tier\":2,\"parent_id\":\"$PARENT_ID\",\"secrets\":$SECRETS_JSON}")
  WID=$(echo "$R" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null)
  WTOK=$(echo "$R" | extract_auth_token)
  [ -n "$WID" ] || fail "$rt workspace create failed: $(printf '%s' "$R" | head -c 300)"
@@ -476,7 +476,19 @@ wait_workspaces_online_routable() {
 # All empty → '{}' (workspace will fail at first turn with an
 # expected, actionable auth error rather than masking the test).
 SECRETS_JSON='{}'
-if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+# Platform-managed path (E2E_LLM_PATH=platform) — the moonshot/kimi
+# NOT_CONFIGURED regression (RFC#340 Fix A #2187). Molecule owns billing via the
+# CP LLM proxy, so the workspace needs NO tenant key: provision with empty
+# secrets and let the workspace boot purely on (a) the proxy env the control
+# plane injects + (b) the manifest-derived `provider: platform` Fix A stamps into
+# the generated config.yaml. This is the path that booted NOT_CONFIGURED in prod
+# precisely because the BYOK branches below never exercise it. We deliberately
+# skip the key-injection branches so a stray E2E_*_API_KEY in the runner env
+# cannot silently convert this into a BYOK run and mask the regression.
+if [ "${E2E_LLM_PATH:-}" = "platform" ]; then
+  log "    LLM path: PLATFORM-MANAGED (no tenant key; proxy + Fix A provider stamp)"
+  SECRETS_JSON='{}'
+elif [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
  SECRETS_JSON=$(python3 -c "
 import json, os
 k = os.environ['E2E_MINIMAX_API_KEY']
@@ -858,6 +870,24 @@ fi
 if echo "$AGENT_TEXT" | grep -qiE "exceeded your current quota|insufficient_quota"; then
  fail "A2A — PROVIDER QUOTA EXHAUSTED (NOT a platform regression). Operator action: top up MOLECULE_STAGING_OPENAI_API_KEY billing or rotate to a higher-quota org at Settings → Secrets and Variables → Actions. Tracked in #2578. Raw: $AGENT_TEXT"
 fi
+# Empty-completion class — the agent runtime reached the LLM and got a
+# 2xx back, but the assistant turn carried NO text part (empty content,
+# or tool_calls/reasoning-only with no surfaced text), so the runtime
+# returns the literal "Error: message contained no text content." as its
+# reply text. Steps 0-7 passing means the platform is healthy (CP up,
+# tenant provisioned, workspace online + routable, A2A delivery e2e); the
+# break is the configured completion BACKEND returning an empty turn — a
+# model/provider-side regression, NOT a workspace-server or harness bug,
+# and NOT NOT_CONFIGURED (that fails earlier, at boot). Name it explicitly
+# so the canary alert points at the model, not the platform: a generic
+# "error-shaped response" misdirects triage to workspace-server. Observed
+# 2026-06-03/04 across every staging canary on MODEL_SLUG=MiniMax-M2 (the
+# canary default since #2710) — 100% on the parent's first cold turn,
+# identical on main's scheduled synthetic E2E and on PRs (so it is an
+# environmental backend regression, never PR-introduced).
+if echo "$AGENT_TEXT" | grep -qiF "message contained no text content"; then
+  fail "A2A — EMPTY COMPLETION (backend regression, NOT a platform/workspace-server bug). The configured model (MODEL_SLUG=${MODEL_SLUG:-?}) returned a 2xx completion with no text part; the runtime surfaced 'message contained no text content.'. Operator action: check the staging LLM backend / proxy for the canary model (MiniMax-M2 since #2710) — empty assistant turns, not an auth/quota/boot fault. Raw: $AGENT_TEXT"
+fi
 # Generic catch-all — falls through if none of the known regressions hit.
 if echo "$AGENT_TEXT" | grep -qiE "error|exception"; then
  fail "A2A returned an error-shaped response: $AGENT_TEXT"
@@ -1011,7 +1041,7 @@ print(json.dumps({
            'messageId': f'e2e-{uuid.uuid4().hex[:8]}',
            'parts': [{'kind': 'text', 'text': 'Reply with exactly: ok'}],
        },
-        'configuration': {'max_tokens': 4}
+        'configuration': {'max_tokens': 32}
    }
 }))
 ")
@@ -26,11 +26,12 @@ import (
 //     the update cycle — no ssh, no re-provision, no ops toil.
 //
 // Contract (paired with cp-side GET /cp/tenants/config):
-//   Request:  GET {MOLECULE_CP_URL or https://api.moleculesai.app}/cp/tenants/config
-//             Authorization: Bearer <ADMIN_TOKEN>
-//             X-Molecule-Org-Id: <MOLECULE_ORG_ID>
-//   Response: 200 {"MOLECULE_CP_SHARED_SECRET":"…","MOLECULE_CP_URL":"…", …}
-//             401 on bearer mismatch or unknown org
+//
+//	Request:  GET {MOLECULE_CP_URL or https://api.moleculesai.app}/cp/tenants/config
+//	          Authorization: Bearer <ADMIN_TOKEN>
+//	          X-Molecule-Org-Id: <MOLECULE_ORG_ID>
+//	Response: 200 {"MOLECULE_CP_SHARED_SECRET":"…","MOLECULE_CP_URL":"…", …}
+//	          401 on bearer mismatch or unknown org
 //
 // Best-effort: any failure logs and returns — main() keeps booting.
 // Self-hosted deploys without MOLECULE_ORG_ID or ADMIN_TOKEN set
@@ -105,3 +106,53 @@ func refreshEnvFromCP() error {
 	log.Printf("CP env refresh: applied %d values from %s/cp/tenants/config", applied, base)
 	return nil
 }
+
+// requiredLLMEnvVars is the set of LLM proxy env vars a managed SaaS
+// tenant must have populated after refreshEnvFromCP. cp#469 (tenant
+// proxy-env delivery) — guaranteed CP-delivered creds reach the
+// tenant process env on boot. Per Researcher Task #37 / Spec 2 and
+// Task #46 (watch-fail-first test).
+//
+// Key set byte-matched against Researcher's verified emission in
+// controlplane tenant_config.go:140-144 (Researcher REQUEST_CHANGES
+// iterate body, 3987f59c). The four keys below ARE the LLM-proxy
+// subset of the 8 CP-emitted keys; OPENAI_BASE_URL / OPENAI_API_KEY /
+// ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY are out of scope for cp#469
+// (different feature surfaces — direct-to-provider fallbacks, not
+// the proxy). v2 fix: MOLECULE_LLM_USAGE_TOKEN, MOLECULE_LLM_USAGE_URL,
+// MOLECULE_LLM_BASE_URL, MOLECULE_LLM_ANTHROPIC_BASE_URL — note the
+// 4th key is namespaced MOLECULE_LLM_ANTHROPIC_BASE_URL, NOT bare
+// ANTHROPIC_BASE_URL. Bare ANTHROPIC_BASE_URL is a separate CP-emitted
+// key for direct-provider use, not the LLM proxy.
+var requiredLLMEnvVars = []string{
+	"MOLECULE_LLM_USAGE_TOKEN",
+	"MOLECULE_LLM_USAGE_URL", // CRITICAL fix v2: was MOLECULE_LLM_URL in v1
+	"MOLECULE_LLM_BASE_URL",
+	"MOLECULE_LLM_ANTHROPIC_BASE_URL", // CRITICAL fix v3: was ANTHROPIC_BASE_URL in v2 (different key!)
+}
+
+// assertManagedTenantHasLLMEnv verifies that, when running as a
+// managed SaaS tenant (MOLECULE_ORG_ID + ADMIN_TOKEN both set), all
+// required LLM proxy env vars are populated after refreshEnvFromCP.
+//
+// Self-hosted (no orgID/adminToken) is exempt — dev must not be
+// blocked here. Managed tenants with missing LLM keys fail with
+// MISSING_CP_LLM_ENV so they do not silently boot with broken proxy
+// creds. Caller in main.go decides whether to log and continue or
+// log.Fatalf depending on deployment context.
+func assertManagedTenantHasLLMEnv() error {
+	if os.Getenv("MOLECULE_ORG_ID") == "" || os.Getenv("ADMIN_TOKEN") == "" {
+		// Self-hosted dev / not yet provisioned — not a managed tenant.
+		return nil
+	}
+	var missing []string
+	for _, k := range requiredLLMEnvVars {
+		if os.Getenv(k) == "" {
+			missing = append(missing, k)
+		}
+	}
+	if len(missing) > 0 {
+		return fmt.Errorf("MISSING_CP_LLM_ENV: required LLM proxy keys not set after refreshEnvFromCP: %v", missing)
+	}
+	return nil
+}
@@ -5,6 +5,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"os"
+	"strings"
 	"testing"
 )

@@ -59,6 +60,138 @@ func TestRefreshEnvFromCP_AppliesCPResponse(t *testing.T) {
 	}
 }

+// TestRefreshEnvFromCP_ManagedTenantRequiresLLMKeys: watch-fail-first
+// per Researcher Task #46. When running as a managed tenant
+// (MOLECULE_ORG_ID + ADMIN_TOKEN set), missing LLM proxy env vars
+// after refreshEnvFromCP MUST surface as MISSING_CP_LLM_ENV, not be
+// silently accepted. Without this guard, a CP that loses its LLM
+// creds (e.g. during an incident) would let a tenant boot and then
+// fail later at first LLM call — worse than a loud refusal here.
+func TestRefreshEnvFromCP_ManagedTenantRequiresLLMKeys(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Stub CP returns a CP response WITHOUT any of the required
+		// LLM keys — simulates the failure mode where the CP side
+		// dropped or never had the LLM creds for this org.
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"MOLECULE_CP_SHARED_SECRET":"x","MOLECULE_CP_URL":"https://api.moleculesai.app"}`)
+	}))
+	defer srv.Close()
+
+	t.Setenv("MOLECULE_ORG_ID", "org-managed-1")
+	t.Setenv("ADMIN_TOKEN", "admin-tok")
+	t.Setenv("MOLECULE_CP_URL", srv.URL)
+	// Clear all LLM keys to simulate the boot-without-LLM-env failure mode.
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "")
+	t.Setenv("MOLECULE_LLM_USAGE_URL", "")
+	t.Setenv("MOLECULE_LLM_BASE_URL", "")
+	t.Setenv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "")
+
+	// refreshEnvFromCP itself should succeed — CP is reachable, returned 200.
+	if err := refreshEnvFromCP(); err != nil {
+		t.Fatalf("refreshEnvFromCP: %v", err)
+	}
+	// The boot assertion must catch the missing LLM keys.
+	err := assertManagedTenantHasLLMEnv()
+	if err == nil {
+		t.Fatal("expected MISSING_CP_LLM_ENV error for managed tenant without LLM keys, got nil")
+	}
+	if !strings.Contains(err.Error(), "MISSING_CP_LLM_ENV") {
+		t.Errorf("expected error to contain MISSING_CP_LLM_ENV, got: %v", err)
+	}
+}
+
+// TestRefreshEnvFromCP_ManagedTenantHappyPath: when the CP returns
+// all 4 LLM-proxy keys, the gate must PASS — no MISSING_CP_LLM_ENV
+// for a properly-configured managed tenant. Watch-fail counterpart
+// to TestRefreshEnvFromCP_ManagedTenantRequiresLLMKeys: if THIS test
+// ever fires MISSING_CP_LLM_ENV on the byte-correct key set, the
+// requiredLLMEnvVars list has drifted from the CP emission again.
+// Per Researcher REQUEST_CHANGES TEST ADEQUACY note.
+func TestRefreshEnvFromCP_ManagedTenantHappyPath(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// Return ALL 4 LLM-proxy keys — names byte-matched to
+		// tenant_config.go:140-144 CP emission.
+		fmt.Fprint(w, `{"MOLECULE_LLM_USAGE_TOKEN":"tok-1","MOLECULE_LLM_USAGE_URL":"https://llm.example.com/usage","MOLECULE_LLM_BASE_URL":"https://llm.example.com","MOLECULE_LLM_ANTHROPIC_BASE_URL":"https://llm.example.com/anthropic"}`)
+	}))
+	defer srv.Close()
+
+	t.Setenv("MOLECULE_ORG_ID", "org-managed-happy")
+	t.Setenv("ADMIN_TOKEN", "admin-tok")
+	t.Setenv("MOLECULE_CP_URL", srv.URL)
+	// Pre-clear so we can verify the refresh actually populated them.
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "")
+	t.Setenv("MOLECULE_LLM_USAGE_URL", "")
+	t.Setenv("MOLECULE_LLM_BASE_URL", "")
+	t.Setenv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "")
+
+	if err := refreshEnvFromCP(); err != nil {
+		t.Fatalf("refreshEnvFromCP: %v", err)
+	}
+	// Sanity: refresh actually applied the keys.
+	if got := os.Getenv("MOLECULE_LLM_USAGE_TOKEN"); got != "tok-1" {
+		t.Errorf("refresh did not apply USAGE_TOKEN: got %q", got)
+	}
+	// The boot assertion must pass — no MISSING_CP_LLM_ENV.
+	if err := assertManagedTenantHasLLMEnv(); err != nil {
+		t.Errorf("managed happy path must not MISSING_CP_LLM_ENV, got: %v", err)
+	}
+}
+
+// TestRefreshEnvFromCP_ManagedTenantPartialEnv: when the CP returns
+// 3 of 4 LLM-proxy keys (one missing), the gate must STILL catch it
+// and the error must name the missing key. Per Researcher
+// REQUEST_CHANGES TEST ADEQUACY note — partial-env coverage is
+// critical because the production failure mode is usually "one
+// key dropped" not "all keys dropped".
+func TestRefreshEnvFromCP_ManagedTenantPartialEnv(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// 3 of 4 — MOLECULE_LLM_ANTHROPIC_BASE_URL is missing.
+		fmt.Fprint(w, `{"MOLECULE_LLM_USAGE_TOKEN":"tok-1","MOLECULE_LLM_USAGE_URL":"https://llm.example.com/usage","MOLECULE_LLM_BASE_URL":"https://llm.example.com"}`)
+	}))
+	defer srv.Close()
+
+	t.Setenv("MOLECULE_ORG_ID", "org-managed-partial")
+	t.Setenv("ADMIN_TOKEN", "admin-tok")
+	t.Setenv("MOLECULE_CP_URL", srv.URL)
+	// Pre-clear all 4 so the 3 that come back from CP are the only
+	// ones set; the 4th (MOLECULE_LLM_ANTHROPIC_BASE_URL) stays empty.
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "")
+	t.Setenv("MOLECULE_LLM_USAGE_URL", "")
+	t.Setenv("MOLECULE_LLM_BASE_URL", "")
+	t.Setenv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "")
+
+	if err := refreshEnvFromCP(); err != nil {
+		t.Fatalf("refreshEnvFromCP: %v", err)
+	}
+	err := assertManagedTenantHasLLMEnv()
+	if err == nil {
+		t.Fatal("expected MISSING_CP_LLM_ENV for partial env (3 of 4 keys), got nil")
+	}
+	if !strings.Contains(err.Error(), "MISSING_CP_LLM_ENV") {
+		t.Errorf("expected error to contain MISSING_CP_LLM_ENV, got: %v", err)
+	}
+	if !strings.Contains(err.Error(), "MOLECULE_LLM_ANTHROPIC_BASE_URL") {
+		t.Errorf("expected error to name the missing key MOLECULE_LLM_ANTHROPIC_BASE_URL, got: %v", err)
+	}
+}
+
+// TestAssertManagedTenantHasLLMEnv_NotManagedIsNoop: self-hosted
+// (no orgID/adminToken) must NOT block on missing LLM keys — dev
+// ergonomics matter and the assertion's contract is "managed only".
+func TestAssertManagedTenantHasLLMEnv_NotManagedIsNoop(t *testing.T) {
+	t.Setenv("MOLECULE_ORG_ID", "")
+	t.Setenv("ADMIN_TOKEN", "")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "")
+	t.Setenv("MOLECULE_LLM_USAGE_URL", "")
+	t.Setenv("MOLECULE_LLM_BASE_URL", "")
+	t.Setenv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "")
+	if err := assertManagedTenantHasLLMEnv(); err != nil {
+		t.Errorf("self-hosted (not managed) must not block, got: %v", err)
+	}
+}
+
 // TestRefreshEnvFromCP_CPUnreachableDoesNotFailBoot: network errors must
 // return non-nil BUT main.go treats that as warn-and-continue. We assert
 // the function returns an error (not a panic) so the caller can log.
@@ -82,6 +82,16 @@ func main() {
 		log.Printf("CP env refresh: %v (continuing with baked-in env)", err)
 	}

+	// Managed-tenant boot assertion (cp#469 — tenant proxy-env delivery).
+	// If we're a managed SaaS tenant (orgID + adminToken set), all required
+	// LLM proxy env vars must be present after refresh. Missing keys block
+	// the tenant from booting with broken LLM creds — silent-fail is worse
+	// than a loud refusal. Self-hosted (no orgID/adminToken) short-circuits
+	// inside the assertion, so this never fires for dev.
+	if err := assertManagedTenantHasLLMEnv(); err != nil {
+		log.Fatalf("Managed tenant boot assertion: %v", err)
+	}
+
 	// Secrets encryption. In MOLECULE_ENV=prod, boot refuses to start
 	// without a valid SECRETS_ENCRYPTION_KEY (fail-secure — Top-5 #5).
 	// In any other environment, missing keys just log a warning and
@@ -349,6 +359,16 @@ func main() {
 		codexauth.StartCodexAuthRefresher(c, db.DB)
 	})

+	// RFC internal#742 Part 2: wire the boot-failure rescue capture into
+	// the provision-timeout sweep's failure verdict. When the sweep flips
+	// a stuck workspace to `failed`, this hook captures a forensic rescue
+	// bundle off the still-running (but boot-failed) EC2 and ships it to
+	// obs/Loki before the control plane reaps the instance. Best-effort +
+	// non-blocking (handlers.BootFailureRescueHook dispatches on its own
+	// goroutine + timeout). The handler-side boot-failure path
+	// (WorkspaceHandler.BootstrapFailed) wires its own capture inline.
+	registry.BootFailureRescueHook = handlers.BootFailureRescueHook
+
 	// Provision-timeout sweep — flips workspaces that have been stuck in
 	// status='provisioning' past the timeout window to 'failed' and emits
 	// WORKSPACE_PROVISION_TIMEOUT. Without this the UI banner is cosmetic
@@ -149,9 +149,11 @@ func markFailed(ctx context.Context, wsID string, broadcaster *events.Broadcaste
 		models.StatusFailed, msg, wsID); dbErr != nil {
 		log.Printf("bundle import: failed to mark workspace %s as failed: %v", wsID, dbErr)
 	}
-	broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceProvisionFailed), wsID, map[string]interface{}{
+	if bcErr := broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceProvisionFailed), wsID, map[string]interface{}{
 		"error": msg,
-	})
+	}); bcErr != nil {
+		log.Printf("bundle import: failed to broadcast provision failed for %s: %v", wsID, bcErr)
+	}
 }

 func nilIfEmpty(s string) interface{} {
@@ -407,12 +407,14 @@ func (m *Manager) HandleInbound(ctx context.Context, ch ChannelRow, msg *Inbound

 	// Broadcast event
 	if m.broadcaster != nil {
-		m.broadcaster.RecordAndBroadcast(ctx, string(events.EventChannelMessage), ch.WorkspaceID, map[string]interface{}{
+		if err := m.broadcaster.RecordAndBroadcast(ctx, string(events.EventChannelMessage), ch.WorkspaceID, map[string]interface{}{
 			"channel_id":   ch.ID,
 			"channel_type": ch.ChannelType,
 			"username":     msg.Username,
 			"direction":    "inbound",
-		})
+		}); err != nil {
+			log.Printf("Channels: failed to broadcast inbound event: %v", err)
+		}
 	}

 	return nil
@@ -453,11 +455,13 @@ func (m *Manager) SendOutbound(ctx context.Context, channelID string, text strin
 	}

 	if m.broadcaster != nil {
-		m.broadcaster.RecordAndBroadcast(ctx, string(events.EventChannelMessage), ch.WorkspaceID, map[string]interface{}{
+		if err := m.broadcaster.RecordAndBroadcast(ctx, string(events.EventChannelMessage), ch.WorkspaceID, map[string]interface{}{
 			"channel_id":   ch.ID,
 			"channel_type": ch.ChannelType,
 			"direction":    "outbound",
-		})
+		}); err != nil {
+			log.Printf("Channels: failed to broadcast outbound event: %v", err)
+		}
 	}

 	return nil
@@ -517,7 +517,9 @@ func (t *TelegramAdapter) StartPolling(ctx context.Context, config map[string]in

 				// Acknowledge the button press (removes loading spinner)
 				ackCfg := tgbotapi.NewCallback(cb.ID, "Received")
-				bot.Send(ackCfg)
+				if _, err := bot.Send(ackCfg); err != nil {
+					log.Printf("telegram: failed to send callback ack: %v", err)
+				}

 				// Update the message to show what was clicked
 				decision := "approved"
@@ -529,7 +531,9 @@ func (t *TelegramAdapter) StartPolling(ctx context.Context, config map[string]in
 					cb.Message.MessageID,
 					cb.Message.Text+"\n\n✅ CEO "+decision,
 				)
-				bot.Send(editMsg)
+				if _, err := bot.Send(editMsg); err != nil {
+					log.Printf("telegram: failed to send edit message: %v", err)
+				}

 				// Route the decision as an inbound message to the agent
 				inbound := &InboundMessage{
@@ -383,23 +383,48 @@ func (h *WorkspaceHandler) logA2ASuccess(ctx context.Context, workspaceID, calle
 	}
 	summary := a2aMethod + " → " + wsNameForLog
 	toolTrace := extractToolTrace(respBody)
-	parent := ctx
-	h.goAsync(func() {
-		logCtx, cancel := context.WithTimeout(context.WithoutCancel(parent), 30*time.Second)
-		defer cancel()
-		LogActivity(logCtx, h.broadcaster, ActivityParams{
-			WorkspaceID:  workspaceID,
-			ActivityType: "a2a_receive",
-			SourceID:     nilIfEmpty(callerID),
-			TargetID:     &workspaceID,
-			Method:       &a2aMethod,
-			Summary:      &summary,
-			RequestBody:  json.RawMessage(body),
-			ResponseBody: json.RawMessage(respBody),
-			ToolTrace:    toolTrace,
-			DurationMs:   &durationMs,
-			Status:       logStatus,
-		})
+
+	// DATA-LOSS FIX (internal#470 / #1347 push-mode sibling): this
+	// a2a_receive row is the ONLY durable record of a push-mode chat
+	// round-trip — request_body carries the user's message, response_body
+	// carries the agent's reply, and chat-history hydration
+	// (messagestore.PostgresMessageStore) reads BOTH back to rebuild the
+	// transcript on canvas reopen / reload. It MUST be written
+	// SYNCHRONOUSLY, before proxyA2ARequest returns and ProxyA2A flushes
+	// the 200 to the canvas — otherwise the canvas sees the reply
+	// acknowledged (and rendered optimistically) while the row is still
+	// racing in a detached goroutine, and a reload (or a workspace-server
+	// restart / deploy / OOM) between the 200 and the goroutine's commit
+	// loses the message permanently on reopen.
+	//
+	// This mirrors the discipline already applied to the poll-mode ingest
+	// path (logA2AReceiveQueued / persistUserMessageAtIngest); the
+	// push-mode counterpart was left async, which the E2E Chat
+	// "history persists across reload" test surfaced as an intermittent
+	// red (the reload out-raced the INSERT).
+	//
+	//   - context.WithoutCancel: a client disconnect on chat-exit (which
+	//     cancels the inbound request ctx) MUST NOT abort this write.
+	//   - SYNCHRONOUS (no goAsync): the row must be durable before the 200.
+	//   - Best-effort: LogActivity logs+swallows INSERT errors internally,
+	//     so a DB hiccup never blocks or fails the user's send — behaviour
+	//     for that one request is never worse than the pre-fix async path.
+	//   - The post-commit ACTIVITY_LOGGED broadcast still fires inside
+	//     LogActivity; the durable row is the truth the canvas re-reads.
+	logCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second)
+	defer cancel()
+	LogActivity(logCtx, h.broadcaster, ActivityParams{
+		WorkspaceID:  workspaceID,
+		ActivityType: "a2a_receive",
+		SourceID:     nilIfEmpty(callerID),
+		TargetID:     &workspaceID,
+		Method:       &a2aMethod,
+		Summary:      &summary,
+		RequestBody:  json.RawMessage(body),
+		ResponseBody: json.RawMessage(respBody),
+		ToolTrace:    toolTrace,
+		DurationMs:   &durationMs,
+		Status:       logStatus,
 	})

 	if callerID == "" && statusCode < 400 {
@@ -60,10 +60,10 @@ func sanitizeErrorDetailForBroadcast(s string) string {
 }

 type ActivityHandler struct {
-	broadcaster *events.Broadcaster
+	broadcaster events.EventEmitter
 }

-func NewActivityHandler(b *events.Broadcaster) *ActivityHandler {
+func NewActivityHandler(b events.EventEmitter) *ActivityHandler {
 	return &ActivityHandler{broadcaster: b}
 }

@@ -54,23 +54,29 @@ func (h *ApprovalsHandler) Create(c *gin.Context) {
 		return
 	}

-	h.broadcaster.RecordAndBroadcast(ctx, string(events.EventApprovalRequested), workspaceID, map[string]interface{}{
+	if err := h.broadcaster.RecordAndBroadcast(ctx, string(events.EventApprovalRequested), workspaceID, map[string]interface{}{
 		"approval_id": approvalID,
 		"action":      body.Action,
 		"reason":      body.Reason,
 		"task_id":     body.TaskID,
-	})
+	}); err != nil {
+		log.Printf("approvals: failed to broadcast approval requested: %v", err)
+	}

 	// Auto-escalate to parent
 	var parentID *string
-	db.DB.QueryRowContext(ctx, `SELECT parent_id FROM workspaces WHERE id = $1`, workspaceID).Scan(&parentID)
+	if err := db.DB.QueryRowContext(ctx, `SELECT parent_id FROM workspaces WHERE id = $1`, workspaceID).Scan(&parentID); err != nil {
+		log.Printf("approvals: failed to lookup parent for escalation: %v", err)
+	}
 	if parentID != nil {
-		h.broadcaster.RecordAndBroadcast(ctx, string(events.EventApprovalEscalated), *parentID, map[string]interface{}{
+		if err := h.broadcaster.RecordAndBroadcast(ctx, string(events.EventApprovalEscalated), *parentID, map[string]interface{}{
 			"approval_id":       approvalID,
 			"from_workspace_id": workspaceID,
 			"action":            body.Action,
 			"reason":            body.Reason,
-		})
+		}); err != nil {
+			log.Printf("approvals: failed to broadcast approval escalated: %v", err)
+		}
 	}

 	c.JSON(http.StatusCreated, gin.H{"approval_id": approvalID, "status": "pending"})
@@ -221,11 +227,13 @@ func (h *ApprovalsHandler) Decide(c *gin.Context) {
 		eventType = "APPROVAL_DENIED"
 	}

-	h.broadcaster.RecordAndBroadcast(ctx, eventType, workspaceID, map[string]interface{}{
+	if err := h.broadcaster.RecordAndBroadcast(ctx, eventType, workspaceID, map[string]interface{}{
 		"approval_id": approvalID,
 		"decision":    body.Decision,
 		"decided_by":  decidedBy,
-	})
+	}); err != nil {
+		log.Printf("approvals: failed to broadcast approval decision: %v", err)
+	}

 	c.JSON(http.StatusOK, gin.H{"status": body.Decision, "approval_id": approvalID})
 }
@@ -196,10 +196,15 @@ func resolveWorkspaceForwardCreds(c *gin.Context, ctx context.Context, workspace
 		c.JSON(http.StatusServiceUnavailable, gin.H{"error": "workspace url not registered yet"})
 		return "", "", false
 	}
-	// Trust note: workspaces.url passes validateAgentURL at /registry/
-	// register write time, blocking SSRF-shaped URLs. We rely on that
-	// upstream gate rather than re-validating here. Tracked at #2316
-	// for follow-up: forward-time re-validation as defense-in-depth.
+	// Defense-in-depth for #2316: workspaces.url is validated at
+	// registration time, but the DB row can be stale/tampered and the
+	// SSRF policy can tighten. Re-validate immediately before attaching
+	// the inbound secret to an outbound forward.
+	if err := isSafeURL(wsURL); err != nil {
+		log.Printf("chat_files %s: unsafe workspace URL for %s rejected: %v", op, workspaceID, err)
+		c.JSON(http.StatusBadRequest, gin.H{"error": "workspace URL not allowed"})
+		return "", "", false
+	}

 	secret, healed, err := readOrLazyHealInboundSecret(ctx, workspaceID, "chat_files "+op)
 	if err != nil {
@@ -414,6 +414,56 @@ func TestChatUpload_WorkspaceUnreachable(t *testing.T) {
 	}
 }

+func TestChatUpload_RejectsMetadataWorkspaceURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	restore := setSSRFCheckForTest(true)
+	t.Cleanup(restore)
+
+	wsID := "00000000-0000-0000-0000-000000000047"
+	expectURL(mock, wsID, "http://169.254.169.254/latest/meta-data")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil, nil))
+	body, ct := uploadFixture(t)
+	c, w := makeUploadRequest(t, wsID, body, ct)
+	h.Upload(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400 for metadata workspace URL, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "workspace URL not allowed") {
+		t.Errorf("expected unsafe URL error, got: %s", w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestChatUpload_RejectsNonHTTPWorkspaceURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	restore := setSSRFCheckForTest(true)
+	t.Cleanup(restore)
+
+	wsID := "00000000-0000-0000-0000-000000000048"
+	expectURL(mock, wsID, "file:///etc/passwd")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil, nil))
+	body, ct := uploadFixture(t)
+	c, w := makeUploadRequest(t, wsID, body, ct)
+	h.Upload(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400 for non-HTTP workspace URL, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "workspace URL not allowed") {
+		t.Errorf("expected unsafe URL error, got: %s", w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
 // TestChatUpload_BodyUnderCap_Forwards pins the lower edge of the new
 // 100 MB body cap (CTO 2026-05-19 directive on forensic a99ab0a1).
 // A multipart payload comfortably under the cap must reach the
@@ -646,6 +696,54 @@ func TestChatDownload_NoInboundSecret_LazyHealFailure(t *testing.T) {
 	}
 }

+func TestChatDownload_RejectsMetadataWorkspaceURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	restore := setSSRFCheckForTest(true)
+	t.Cleanup(restore)
+
+	wsID := "00000000-0000-0000-0000-000000000054"
+	expectURL(mock, wsID, "http://169.254.169.254/latest/meta-data")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil, nil))
+	c, w := makeDownloadRequest(t, wsID, "/workspace/foo.txt")
+	h.Download(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400 for metadata workspace URL, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "workspace URL not allowed") {
+		t.Errorf("expected unsafe URL error, got: %s", w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestChatDownload_RejectsNonHTTPWorkspaceURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	restore := setSSRFCheckForTest(true)
+	t.Cleanup(restore)
+
+	wsID := "00000000-0000-0000-0000-000000000055"
+	expectURL(mock, wsID, "file:///etc/passwd")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil, nil))
+	c, w := makeDownloadRequest(t, wsID, "/workspace/foo.txt")
+	h.Download(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400 for non-HTTP workspace URL, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "workspace URL not allowed") {
+		t.Errorf("expected unsafe URL error, got: %s", w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
 func TestChatDownload_ForwardsToWorkspace_HappyPath(t *testing.T) {
 	mock := setupTestDB(t)
 	setupTestRedis(t)
@@ -102,10 +102,10 @@ func pushDelegationResultToInbox(ctx context.Context, sourceID, delegationID, st
 // and the A2A request runs in the background.
 type DelegationHandler struct {
 	workspace   *WorkspaceHandler
-	broadcaster *events.Broadcaster
+	broadcaster events.EventEmitter
 }

-func NewDelegationHandler(wh *WorkspaceHandler, b *events.Broadcaster) *DelegationHandler {
+func NewDelegationHandler(wh *WorkspaceHandler, b events.EventEmitter) *DelegationHandler {
 	return &DelegationHandler{workspace: wh, broadcaster: b}
 }

@@ -36,7 +36,6 @@ package handlers
 import (
 	"context"
 	"database/sql"
-	"os"
 	"strings"
 	"testing"
 	"time"
@@ -57,10 +56,7 @@ import (
 // directly rather than going through the package global.
 func integrationDB(t *testing.T) *sql.DB {
 	t.Helper()
-	url := os.Getenv("INTEGRATION_DB_URL")
-	if url == "" {
-		t.Skip("INTEGRATION_DB_URL not set; skipping (local devs: see file header)")
-	}
+	url := requireIntegrationDBURL(t)
 	conn, err := sql.Open("postgres", url)
 	if err != nil {
 		t.Fatalf("open: %v", err)
@@ -0,0 +1,40 @@
+//go:build integration
+// +build integration
+
+// integration_helper_test.go — shared preflight for handler Postgres
+// integration tests. Extracted so the fail-open/skip logic is in ONE place
+// and can be tightened without editing every integration test file.
+//
+// See delegation_ledger_integration_test.go for the docker-postgres setup
+// incantation used by local devs.
+
+package handlers
+
+import (
+	"os"
+	"testing"
+)
+
+// requireIntegrationDBURL returns $INTEGRATION_DB_URL.
+//
+// In CI (CI, GITHUB_ACTIONS, or GITEA_ACTIONS env var is non-empty), an
+// empty URL is a fatal error — it means the workflow failed to export the
+// variable (postgres container did not start, bridge IP resolution failed,
+// or a regression in the workflow YAML). t.Fatalf keeps the test red so the
+// failure is visible; t.Skip would silently pass and mask the defect.
+//
+// Locally (none of the three CI markers set), an empty URL skips the test
+// so devs can run `go test ./...` without booting a Postgres container.
+func requireIntegrationDBURL(t *testing.T) string {
+	t.Helper()
+	url := os.Getenv("INTEGRATION_DB_URL")
+	if url == "" {
+		if os.Getenv("CI") != "" ||
+			os.Getenv("GITHUB_ACTIONS") != "" ||
+			os.Getenv("GITEA_ACTIONS") != "" {
+			t.Fatalf("INTEGRATION_DB_URL required in CI handler integration tests — check workflow env export")
+		}
+		t.Skip("INTEGRATION_DB_URL not set; skipping (local devs: see file header)")
+	}
+	return url
+}
@@ -372,3 +372,78 @@ func TestApplyPlatformManagedLLMEnv_WorkspaceOriginCredExemptFromStrip(t *testin
 		t.Errorf("sqlmock expectations: %v", err)
 	}
 }
+
+// TestApplyPlatformManagedLLMEnv_MissingProxyEnvFailClosed is the #2162
+// regression guard. A platform-managed workspace whose CP proxy env is absent
+// must NOT start credential-less. The empty-proxy path must return
+// HasUsableLLMCred=false so the caller aborts with MISSING_PLATFORM_PROXY.
+//
+// Mutation: revert the early-return from HasUsableLLMCred=false to true
+// → workspace starts with zero credential → "container started but never
+// called /registry/register" (600s provision-timeout sweep) → this test RED.
+func TestApplyPlatformManagedLLMEnv_MissingProxyEnvFailClosed(t *testing.T) {
+	ctx := context.Background()
+	const wsID = "29b95be9-811e-4857-be36-1dafdbf4f697" // adk-demo failure workspace
+
+	mock := setupTestDB(t)
+	expectOverrideQuery(mock, wsID, "")
+
+	// No proxy env present — simulates the boot-race / misconfig path.
+	envVars := map[string]string{}
+	res := applyPlatformManagedLLMEnv(ctx, envVars, wsID, "claude-code", "moonshot/kimi-k2.6", nil)
+
+	if res.ResolvedMode != LLMBillingModePlatformManaged {
+		t.Fatalf("platform-managed model must stay platform_managed, got %q (source=%s)", res.ResolvedMode, res.Source)
+	}
+	// THE FIX: must NOT report usable credential when none was injected.
+	if res.HasUsableLLMCred {
+		t.Fatalf("empty proxy env → HasUsableLLMCred must be false (fail-closed), got true — the #2162 dark-wedge class")
+	}
+	// No credential env must be present.
+	if _, present := envVars["ANTHROPIC_API_KEY"]; present {
+		t.Errorf("empty proxy env must NOT inject ANTHROPIC_API_KEY")
+	}
+	if _, present := envVars["MOLECULE_LLM_USAGE_TOKEN"]; present {
+		t.Errorf("empty proxy env must NOT inject MOLECULE_LLM_USAGE_TOKEN")
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations: %v", err)
+	}
+}
+
+// TestApplyPlatformManagedLLMEnv_ProxyEnvPresentInjectsCredential is the
+// positive-path pair to the #2162 regression guard: when the CP proxy env IS
+// present, the platform-managed path must inject ANTHROPIC_API_KEY +
+// ANTHROPIC_BASE_URL for an Anthropic-native runtime and report
+// HasUsableLLMCred=true.
+func TestApplyPlatformManagedLLMEnv_ProxyEnvPresentInjectsCredential(t *testing.T) {
+	ctx := context.Background()
+	const wsID = "aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee"
+
+	mock := setupTestDB(t)
+	expectOverrideQuery(mock, wsID, "")
+
+	envVars := map[string]string{}
+	// Simulate the CP proxy env being present (as it is in production).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.moleculesai.app/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "https://api.moleculesai.app/api/v1/internal/llm/anthropic/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "PLATFORM-PROXY-TOKEN")
+
+	res := applyPlatformManagedLLMEnv(ctx, envVars, wsID, "claude-code", "moonshot/kimi-k2.6", nil)
+
+	if res.ResolvedMode != LLMBillingModePlatformManaged {
+		t.Fatalf("expected platform_managed, got %q", res.ResolvedMode)
+	}
+	if !res.HasUsableLLMCred {
+		t.Fatalf("proxy env present → HasUsableLLMCred must be true, got false")
+	}
+	if envVars["ANTHROPIC_API_KEY"] != "PLATFORM-PROXY-TOKEN" {
+		t.Errorf("ANTHROPIC_API_KEY must be injected with the platform proxy token; got %q", envVars["ANTHROPIC_API_KEY"])
+	}
+	if envVars["ANTHROPIC_BASE_URL"] != "https://api.moleculesai.app/api/v1/internal/llm/anthropic/v1" {
+		t.Errorf("ANTHROPIC_BASE_URL must be injected with the platform anthropic proxy; got %q", envVars["ANTHROPIC_BASE_URL"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations: %v", err)
+	}
+}
@@ -17,6 +17,7 @@ package handlers

 import (
 	"fmt"
+	"sort"
 	"strings"
 )

@@ -55,3 +56,95 @@ func validateRegisteredModelForRuntime(runtime, model string) (bool, string) {
 		"model %q is not a registered model for runtime %q; pick one of the runtime's registered models (provider-registry SSOT, internal#718)",
 		model, runtime)
 }
+
+// validateDerivedProviderInRegistry (issue #2172) is the provider-side companion
+// to validateRegisteredModelForRuntime. The model-side check asks "is this
+// (runtime, model) in the registry?"; the provider-side check asks "is the
+// provider this model DERIVES to — the same one the adapter will resolve at
+// boot — a known provider in providers.yaml?"
+//
+// Live trigger (adk-demo Assistant, 2026-06-03): workspace config
+// `model=moonshot/kimi-k2.6` (claude-code) → adapter derives `provider=moonshot`
+// → `ValueError: provider=moonshot not in providers registry` at BOOT. The
+// save was accepted (no validation at the API boundary), and the failure only
+// surfaced when the agent tried to register. CI never saw it. The drift gate
+// (RFC#580) validates TEMPLATES against the registry, NOT per-workspace
+// configs; the existing model-side check rejects a model the runtime doesn't
+// own but says nothing about the DERIVED provider's registry membership.
+//
+// Returns:
+//
+//	(true,  "")     — pass: model is empty (MODEL_REQUIRED owns it), the
+//	                  runtime is not in the registry (fail-open for
+//	                  federated / non-first-party runtimes — mirror of the
+//	                  model-side check's federation contract), the registry
+//	                  failed to load (build-time gate owns it), OR the
+//	                  derived provider name is a known provider in the
+//	                  registry's `providers:` list.
+//	(false, reason) — reject: a known (runtime, model) pair derives to a
+//	                  provider name absent from the providers list. This is
+//	                  the structural class the adk-demo boot failure belongs
+//	                  to — the registry's `runtimes:` block references a
+//	                  provider not declared in `providers:`, which by
+//	                  construction is a registry-data bug. Catching it at
+//	                  config-SAVE keeps it out of the agent-boot path.
+//
+// Defense-in-depth: by construction, a model in a runtime's native provider set
+// has a provider that IS in the catalog (the runtime ref names a provider from
+// the providers list). So the rejection path is primarily a registry-consistency
+// guard. The real value is the FAIL-LOUD semantics — any future drift between
+// `providers:` and `runtimes:` fails the create call with a clear pointer to
+// the missing provider, instead of silently wedging the agent at boot.
+func validateDerivedProviderInRegistry(runtime, model string) (bool, string) {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return true, "" // MODEL_REQUIRED owns this.
+	}
+	m, err := providerRegistry()
+	if err != nil || m == nil {
+		// Registry unavailable (build-time defect the gates catch). Fail open —
+		// do not block create on a registry-load failure.
+		return true, ""
+	}
+	// DeriveProvider is fail-closed for unknown runtimes. Mirror the
+	// model-side check's federation contract: a runtime the registry does
+	// NOT know (langgraph / external / kimi / mock / federated) is allowed
+	// to pass through. DeriveProvider's `unknown runtime` error IS that
+	// signal — treat it as fail-open, identical to ModelsForRuntime's
+	// not-found behavior above.
+	p, err := m.DeriveProvider(runtime, model, nil)
+	if err != nil {
+		// Either the runtime is unknown (fail-open by contract) OR the model
+		// is not native to the runtime (the model-side validator already
+		// rejected this — DeriveProvider's error here means
+		// validateRegisteredModelForRuntime should have caught it. Don't
+		// double-reject: pass through and let the model-side response own
+		// the message).
+		return true, ""
+	}
+	// Defense-in-depth: confirm the DERIVED provider is a known entry in the
+	// providers list. By construction it should be (DeriveProvider only
+	// returns a Provider that was looked up by name from `providers:`), but
+	// a future federation merge could introduce a runtime ref pointing at a
+	// contributed provider absent from the core catalog. Reject loudly here
+	// rather than letting the save reach the agent-boot path and wedge with
+	// "provider=X not in providers registry" (the original adk-demo class).
+	for _, candidate := range m.Providers {
+		if candidate.Name == p.Name {
+			return true, ""
+		}
+	}
+	// Build a sorted, comma-separated list of valid provider names so the
+	// operator/caller sees the actionable list (the boot-time error message
+	// the adk-demo class produced does NOT include this — the fix is to
+	// surface it at the API boundary, where the caller can fix the request
+	// without a stuck workspace + operator page).
+	valid := make([]string, 0, len(m.Providers))
+	for _, c := range m.Providers {
+		valid = append(valid, c.Name)
+	}
+	sort.Strings(valid)
+	return false, fmt.Sprintf(
+		"derived provider %q (for model %q on runtime %q) is not in the providers registry; pick a model whose derived provider is one of: %s",
+		p.Name, model, runtime, strings.Join(valid, ", "))
+}
@@ -6,8 +6,17 @@ package handlers
 // fail OPEN (allow) for a runtime the registry doesn't know yet (federation /
 // langgraph/etc. not in the first-party registry) so the existing knownRuntimes
 // gate stays authoritative there.
+//
+// TestValidateDerivedProviderInRegistry (issue #2172) is the provider-side
+// companion: once the model-side check passes, confirm the DERIVED provider
+// (the one the adapter will resolve at boot) is a known provider in
+// providers.yaml. Catches the adk-demo "provider=X not in providers registry"
+// class at config-SAVE time instead of letting it wedge the agent at boot.

-import "testing"
+import (
+	"strings"
+	"testing"
+)

 func TestValidateRegisteredModelForRuntime(t *testing.T) {
 	type tc struct {
@@ -80,3 +89,163 @@ func TestValidateRegisteredModelForRuntime(t *testing.T) {
 		})
 	}
 }
+
+func TestValidateDerivedProviderInRegistry(t *testing.T) {
+	type tc struct {
+		name    string
+		runtime string
+		model   string
+		wantOK  bool
+		// wantReasonContains: a substring the rejection reason must include
+		// (skipped for OK cases). Pins the actionable list / derivation pointer
+		// so the caller knows which provider was missing and what the valid
+		// set looks like — this is the fix that distinguishes the new gate
+		// from the boot-time "provider=X not in providers registry" string
+		// it replaces.
+		wantReasonContains string
+	}
+	cases := []tc{
+		// PASS — every native (runtime, model) in the catalog derives to a
+		// provider that IS in the providers list. These are the live corpus
+		// entries; the test pins the registry-consistency invariant.
+		{
+			name:   "claude_code_anthropic_api_native",
+			runtime: "claude-code",
+			model:   "claude-sonnet-4-6",
+			wantOK: true,
+		},
+		{
+			name:   "claude_code_kimi_coding_native",
+			runtime: "claude-code",
+			model:   "kimi-for-coding",
+			wantOK: true,
+		},
+		{
+			name:   "claude_code_minimax_native",
+			runtime: "claude-code",
+			model:   "MiniMax-M2.7",
+			wantOK: true,
+		},
+		{
+			name:   "claude_code_platform_namespaced",
+			runtime: "claude-code",
+			model:   "moonshot/kimi-k2.6",
+			wantOK: true,
+		},
+		{
+			name:   "codex_openai_subscription_default_arm",
+			runtime: "codex",
+			model:   "gpt-5.5",
+			wantOK: true,
+		},
+		{
+			name:   "codex_platform_namespaced",
+			runtime: "codex",
+			model:   "openai/gpt-5.4-mini",
+			wantOK: true,
+		},
+		{
+			name:   "hermes_kimi_coding",
+			runtime: "hermes",
+			model:   "kimi-coding/kimi-k2",
+			wantOK: true,
+		},
+		{
+			name:   "hermes_platform_namespaced",
+			runtime: "hermes",
+			model:   "moonshot/kimi-k2.6",
+			wantOK: true,
+		},
+		{
+			name:   "openclaw_kimi_coding",
+			runtime: "openclaw",
+			model:   "moonshot:kimi-k2.6",
+			wantOK: true,
+		},
+		// FAIL — model-side validator catches this, but the provider-side
+		// gate is called AFTER it in Create and inherits the fail-open
+		// contract for "model is not native to runtime" (DeriveProvider
+		// errors → allow, letting the model-side response own the message).
+		// This is the deliberate "don't double-reject" decision.
+		{
+			name:   "unregistered_model_pass_through_to_model_side",
+			runtime: "claude-code",
+			model:   "totally-made-up-model-xyz",
+			wantOK: true, // pass-through: model-side validator owns the rejection
+		},
+		// Federation contract — mirror of the model-side test above.
+		{
+			name:   "langgraph_runtime_failopen",
+			runtime: "langgraph",
+			model:   "anything-goes",
+			wantOK: true,
+		},
+		{
+			name:   "external_runtime_failopen",
+			runtime: "external",
+			model:   "whatever",
+			wantOK: true,
+		},
+		// Empty model — MODEL_REQUIRED owns it; allow.
+		{
+			name:   "empty_model_allowed_other_gate_owns_it",
+			runtime: "claude-code",
+			model:   "",
+			wantOK: true,
+		},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			ok, why := validateDerivedProviderInRegistry(c.runtime, c.model)
+			if ok != c.wantOK {
+				t.Errorf("validateDerivedProviderInRegistry(%q,%q) ok=%v want %v (reason=%q)",
+					c.runtime, c.model, ok, c.wantOK, why)
+			}
+			if !c.wantOK && c.wantReasonContains != "" && !strings.Contains(why, c.wantReasonContains) {
+				t.Errorf("rejection reason missing %q: got %q", c.wantReasonContains, why)
+			}
+		})
+	}
+}
+
+// TestRegistryConsistency_AllNativeModelsDeriveToKnownProvider walks every
+// (runtime, model) pair in the registry's native model sets and asserts each
+// one derives to a provider that IS in the providers list. This is the
+// static regression gate the issue calls for ("a CI test fails if any shipped
+// demo/template config references an unregistered provider") — generalized
+// to the catalog as a whole: if anyone edits providers.yaml such that a
+// `runtimes:` block names a provider absent from `providers:`, this test
+// fires before the bad config can reach a customer workspace.
+//
+// By construction this invariant should always hold (DeriveProvider only
+// returns a Provider that was looked up by name from `providers:`), so the
+// test primarily guards against future federation merges that introduce a
+// runtime ref pointing at a contributed provider absent from the core
+// catalog — exactly the failure shape the adk-demo Assistant wedge
+// belongs to.
+func TestRegistryConsistency_AllNativeModelsDeriveToKnownProvider(t *testing.T) {
+	m, err := providerRegistry()
+	if err != nil || m == nil {
+		t.Skipf("providerRegistry unavailable in test env (err=%v); skipping consistency walk", err)
+	}
+	providerNames := make(map[string]struct{}, len(m.Providers))
+	for _, p := range m.Providers {
+		providerNames[p.Name] = struct{}{}
+	}
+	for runtimeName, runtime := range m.Runtimes {
+		for _, ref := range runtime.Providers {
+			for _, modelID := range ref.Models {
+				p, err := m.DeriveProvider(runtimeName, modelID, nil)
+				if err != nil {
+					t.Errorf("catalog invariant broken: runtime=%q model=%q failed DeriveProvider: %v",
+						runtimeName, modelID, err)
+					continue
+				}
+				if _, ok := providerNames[p.Name]; !ok {
+					t.Errorf("catalog invariant broken: runtime=%q model=%q derives to provider %q which is not in the providers list (refs=%q)",
+						runtimeName, modelID, p.Name, ref.Name)
+				}
+			}
+		}
+	}
+}
@@ -43,7 +43,6 @@ package handlers
 import (
 	"context"
 	"database/sql"
-	"os"
 	"strings"
 	"testing"
 	"time"
@@ -63,10 +62,7 @@ import (
 // but kept separate so each table's wipe step is local to its tests.
 func integrationDB_PendingUploads(t *testing.T) *sql.DB {
 	t.Helper()
-	url := os.Getenv("INTEGRATION_DB_URL")
-	if url == "" {
-		t.Skip("INTEGRATION_DB_URL not set; skipping (local devs: see file header)")
-	}
+	url := requireIntegrationDBURL(t)
 	conn, err := sql.Open("postgres", url)
 	if err != nil {
 		t.Fatalf("open: %v", err)
@@ -171,11 +171,9 @@ func (h *PluginsHandler) uninstallViaDocker(ctx context.Context, c *gin.Context,
 			log.Printf("Plugin uninstall: skipping invalid skill name %q in %s: %v", skill, pluginName, err)
 			continue
 		}
-		if _, rmErr := h.execAsRoot(ctx, containerName, []string{
+		_, _ = h.execAsRoot(ctx, containerName, []string{
 			"rm", "-rf", "/configs/skills/" + skill,
-		}); rmErr != nil {
-			log.Printf("Plugin uninstall: failed to remove skill %s from %s: %v", skill, workspaceID, rmErr)
-		}
+		})
 	}

 	// 3. Delete the plugin directory itself (as root to handle file ownership).
@@ -417,9 +417,7 @@ func (h *PluginsHandler) stripPluginMarkersFromMemory(ctx context.Context, conta
 		`awk 'BEGIN{skip=0; blanks=0} /^%s/{skip=1; blanks=0; next} skip==1 && /^[[:space:]]*$/{blanks++; if(blanks>=2){skip=0; print; next} next} /^# Plugin: /{if(skip==1)skip=0} skip==1{next} {print}' /configs/CLAUDE.md > /tmp/claude.new && mv /tmp/claude.new /configs/CLAUDE.md`,
 		regexpEscapeForAwk(marker),
 	)
-	if _, awkErr := h.execAsRoot(ctx, containerName, []string{"bash", "-c", script}); awkErr != nil {
-		log.Printf("Plugin uninstall: failed to strip markers from CLAUDE.md for %s in %s: %v", pluginName, workspaceID, awkErr)
-	}
+	_, _ = h.execAsRoot(ctx, containerName, []string{"bash", "-c", script})
 }

 // regexpEscapeForAwk escapes characters that have special meaning inside an
@@ -0,0 +1,21 @@
+package handlers
+
+import (
+	"testing"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/providers"
+)
+
+// Proper-SSOT (task #65): required_env is DERIVED from the resolved provider's
+// serving classification (IsPlatform), not hand-authored — platform injects
+// creds server-side (none required), BYOK requires its auth_env.
+func TestRequiredEnvForRegistryProvider(t *testing.T) {
+	if got := requiredEnvForRegistryProvider(providers.Provider{Name: providers.PlatformProviderName}); got != nil {
+		t.Errorf("platform provider requiredEnv = %v; want nil (creds injected server-side)", got)
+	}
+	byok := providers.Provider{Name: "google", AuthEnv: []string{"GEMINI_API_KEY", "GOOGLE_API_KEY"}}
+	got := requiredEnvForRegistryProvider(byok)
+	if len(got) != 2 || got[0] != "GEMINI_API_KEY" {
+		t.Errorf("byok requiredEnv = %v; want its auth_env", got)
+	}
+}
@@ -0,0 +1,168 @@
+package handlers
+
+// rescue_read.go — GET /workspaces/:id/rescue (RFC internal#742 Part 3).
+//
+// Serves the LATEST post-mortem rescue bundle captured for a
+// boot-failed/terminated workspace, so "why won't my agent boot" is
+// answerable WITHOUT a live instance. Powers the future canvas
+// "Why did this fail?" panel.
+//
+// Read-path: the bundle is read from the queryable rescue_bundles table
+// (internal/rescuestore), NOT from obs/Loki. Part 2 ships the bundle via
+// internal/audit (Loki-only); reading from Loki would require obs read
+// creds the tenant deliberately lacks. Part 3 persists the
+// already-redacted bundle on capture and serves it here — see the
+// migration header for the full rationale.
+//
+// Auth/scoping: registered on the WorkspaceAuth-guarded /workspaces/:id
+// group (same gate as /files/* and /exec), so the caller must hold a
+// valid per-workspace or org bearer token for :id. TenantGuard already
+// 404s cross-org requests at the routing layer; on top of that the store
+// read is org-scoped by MOLECULE_ORG_ID, so a row written under a
+// different org is never returned (defense in depth).
+//
+// Redaction: the stored sections were already scrubbed at capture time
+// (Part 2's SAFE-T1201 secret-scan). This handler returns them verbatim
+// — it never re-ships or re-derives secrets.
+
+import (
+	"log"
+	"net/http"
+	"os"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescuestore"
+	"github.com/gin-gonic/gin"
+)
+
+// maxResponseSections bounds how many sections the read response
+// returns. The fixed capture set is small (6), so this is a backstop
+// against a future capture set growth or a hand-written row — keeps the
+// JSON response bounded regardless of what's stored. Per-section content
+// is already clamped at persist time (rescuestore.maxSectionBytes).
+const maxResponseSections = 64
+
+// RescueReadHandler serves GET /workspaces/:id/rescue. The store is
+// injected so tests fake it; production wires a Postgres store over
+// db.DB (see NewRescueReadHandler).
+type RescueReadHandler struct {
+	store rescuestore.Store
+}
+
+// NewRescueReadHandler builds the handler over the package db.DB. db.DB
+// is nil in some unit-test binaries; the handler tolerates that by
+// returning 503 rather than nil-deref (the store guards nil db).
+func NewRescueReadHandler() *RescueReadHandler {
+	return &RescueReadHandler{store: rescuestore.NewPostgres(db.DB)}
+}
+
+// WithStore overrides the store (test seam). Returns the handler for
+// chaining.
+func (h *RescueReadHandler) WithStore(s rescuestore.Store) *RescueReadHandler {
+	h.store = s
+	return h
+}
+
+// rescueSection is one labelled chunk in the read response.
+type rescueSection struct {
+	Name     string `json:"name"`
+	Content  string `json:"content"`
+	Redacted bool   `json:"redacted"`
+}
+
+// rescueReadResponse is the JSON shape returned for a found bundle.
+// `sections` is an ordered array (capture reading order), not a map, so
+// the order config→logs→state→env is preserved for the canvas panel.
+type rescueReadResponse struct {
+	WorkspaceID string          `json:"workspace_id"`
+	CapturedAt  time.Time       `json:"captured_at"`
+	Reason      string          `json:"reason"`
+	InstanceID  string          `json:"instance_id"`
+	Sections    []rescueSection `json:"sections"`
+	// Truncated is true when the stored bundle had more sections than
+	// maxResponseSections and the response was capped.
+	Truncated bool `json:"truncated,omitempty"`
+}
+
+// GetRescue handles GET /workspaces/:id/rescue.
+//
+//	200 — latest rescue bundle for the workspace (org-scoped).
+//	404 — no rescue bundle on file for this workspace (or wrong org).
+//	503 — store/datastore unavailable.
+func (h *RescueReadHandler) GetRescue(c *gin.Context) {
+	workspaceID := c.Param("id")
+	ctx := c.Request.Context()
+
+	if h.store == nil {
+		log.Printf("GetRescue: store not configured for ws=%s", workspaceID)
+		c.JSON(http.StatusServiceUnavailable, gin.H{
+			"error": "rescue store unavailable",
+			"code":  "platform_unavailable",
+		})
+		return
+	}
+
+	// org_id is the tenant's configured org (one tenant = one org).
+	// Fail closed: an empty org_id disables org isolation and must not
+	// reach the store (#2020).
+	orgID := os.Getenv("MOLECULE_ORG_ID")
+	if orgID == "" {
+		log.Printf("GetRescue: missing MOLECULE_ORG_ID for ws=%s", workspaceID)
+		c.JSON(http.StatusServiceUnavailable, gin.H{
+			"error": "rescue org not configured",
+			"code":  "platform_misconfigured",
+		})
+		return
+	}
+
+	stored, err := h.store.GetLatest(ctx, workspaceID, orgID)
+	if err != nil {
+		// Per the Store contract a missing bundle is (nil, nil), NOT an
+		// error — so any error here is a genuine datastore fault → 503,
+		// never a masquerading 404 that would hide an outage.
+		log.Printf("GetRescue: store query failed for ws=%s: %v", workspaceID, err)
+		c.JSON(http.StatusServiceUnavailable, gin.H{
+			"error": "rescue store query failed",
+			"code":  "platform_unavailable",
+		})
+		return
+	}
+	if stored == nil {
+		// No bundle captured (workspace never boot-failed, or its grace
+		// window lapsed). 404 — existence-non-inferring; a workspace in a
+		// sibling org reaches the same 404 via the org filter.
+		c.JSON(http.StatusNotFound, gin.H{"error": "no rescue bundle for this workspace"})
+		return
+	}
+
+	resp := buildRescueResponse(workspaceID, stored)
+	c.JSON(http.StatusOK, resp)
+}
+
+// buildRescueResponse maps a stored bundle to the read response, bounding
+// the section count. Split out so the mapping/limit is unit-testable.
+func buildRescueResponse(workspaceID string, stored *rescuestore.StoredBundle) rescueReadResponse {
+	secs := stored.Bundle.Sections
+	truncated := false
+	if len(secs) > maxResponseSections {
+		secs = secs[:maxResponseSections]
+		truncated = true
+	}
+	out := make([]rescueSection, 0, len(secs))
+	for _, s := range secs {
+		// rescue.Section and rescueSection are field-identical; the
+		// explicit conversion keeps the handler's JSON shape independent
+		// of the leaf package's struct (which could gain non-response
+		// fields later).
+		out = append(out, rescueSection(s))
+	}
+	return rescueReadResponse{
+		WorkspaceID: workspaceID,
+		CapturedAt:  stored.CapturedAt,
+		Reason:      stored.Bundle.Reason,
+		InstanceID:  stored.Bundle.InstanceID,
+		Sections:    out,
+		Truncated:   truncated,
+	}
+}
@@ -0,0 +1,238 @@
+package handlers
+
+// Tests for GET /workspaces/:id/rescue (RFC internal#742 Part 3).
+//
+// These exercise the handler against a FAKE store (no DB) so every path
+// is deterministic without external infra:
+//   - returns the latest bundle in the documented shape
+//   - 404 when no bundle exists for the workspace
+//   - org-scoping: the handler passes the tenant's MOLECULE_ORG_ID to
+//     the store, so a fake that returns nil for a mismatched org proves a
+//     sibling org cannot read another org's bundle
+//   - 503 on a store/datastore error (not a 404 masquerade)
+//   - redaction/shape preserved: stored sections are returned verbatim,
+//     no re-derivation
+//
+// WorkspaceAuth gating itself is covered by the middleware tests; here we
+// invoke the handler directly (the route is registered on the wsAuth
+// group in router.go).
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescuestore"
+	"github.com/gin-gonic/gin"
+)
+
+func init() { gin.SetMode(gin.TestMode) }
+
+// fakeRescueStore records the args it was called with and returns a
+// scripted result. Implements rescuestore.Store.
+type fakeRescueStore struct {
+	// gotWorkspaceID/gotOrgID capture what the handler passed.
+	gotWorkspaceID string
+	gotOrgID       string
+	// ret/err are the scripted GetLatest result.
+	ret *rescuestore.StoredBundle
+	err error
+}
+
+func (f *fakeRescueStore) Persist(_ context.Context, _ rescue.Bundle) error { return nil }
+
+func (f *fakeRescueStore) GetLatest(_ context.Context, workspaceID, orgID string) (*rescuestore.StoredBundle, error) {
+	f.gotWorkspaceID = workspaceID
+	f.gotOrgID = orgID
+	return f.ret, f.err
+}
+
+// doRescueGet runs the handler for ws against the given fake and returns
+// the recorder. orgEnv sets MOLECULE_ORG_ID for the duration.
+func doRescueGet(t *testing.T, ws, orgEnv string, fake *fakeRescueStore) *httptest.ResponseRecorder {
+	t.Helper()
+	t.Setenv("MOLECULE_ORG_ID", orgEnv)
+
+	h := (&RescueReadHandler{}).WithStore(fake)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: ws}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/"+ws+"/rescue", nil)
+	h.GetRescue(c)
+	return w
+}
+
+// sampleStored builds a representative stored bundle with a redacted +
+// a failure-marker section.
+func sampleStored() *rescuestore.StoredBundle {
+	return &rescuestore.StoredBundle{
+		CapturedAt: time.Date(2026, 5, 31, 12, 0, 0, 0, time.UTC),
+		Bundle: rescue.Bundle{
+			WorkspaceID: "ws-1",
+			OrgID:       "org-9",
+			InstanceID:  "i-abc123",
+			Reason:      "provision_timeout_sweep",
+			Sections: []rescue.Section{
+				{Name: "config.yaml", Content: "model: gpt-4\nANTHROPIC_API_KEY=[REDACTED]", Redacted: true},
+				{Name: "docker-ps", Content: "(rescue: section collection failed: ssh blip)", Redacted: false},
+			},
+		},
+	}
+}
+
+// TestGetRescue_ReturnsLatestBundle — happy path: 200 with the full
+// documented shape, sections in order, redaction-preserved.
+func TestGetRescue_ReturnsLatestBundle(t *testing.T) {
+	fake := &fakeRescueStore{ret: sampleStored()}
+	w := doRescueGet(t, "ws-1", "org-9", fake)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200; body=%s", w.Code, w.Body.String())
+	}
+	var resp struct {
+		WorkspaceID string    `json:"workspace_id"`
+		CapturedAt  time.Time `json:"captured_at"`
+		Reason      string    `json:"reason"`
+		InstanceID  string    `json:"instance_id"`
+		Sections    []struct {
+			Name     string `json:"name"`
+			Content  string `json:"content"`
+			Redacted bool   `json:"redacted"`
+		} `json:"sections"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("decode: %v; body=%s", err, w.Body.String())
+	}
+	if resp.WorkspaceID != "ws-1" {
+		t.Errorf("workspace_id = %q, want ws-1", resp.WorkspaceID)
+	}
+	if resp.Reason != "provision_timeout_sweep" {
+		t.Errorf("reason = %q", resp.Reason)
+	}
+	if resp.InstanceID != "i-abc123" {
+		t.Errorf("instance_id = %q", resp.InstanceID)
+	}
+	if !resp.CapturedAt.Equal(time.Date(2026, 5, 31, 12, 0, 0, 0, time.UTC)) {
+		t.Errorf("captured_at = %v", resp.CapturedAt)
+	}
+	if len(resp.Sections) != 2 {
+		t.Fatalf("sections = %d, want 2", len(resp.Sections))
+	}
+	// Order preserved: config first, docker-ps second.
+	if resp.Sections[0].Name != "config.yaml" || resp.Sections[1].Name != "docker-ps" {
+		t.Errorf("section order wrong: %q, %q", resp.Sections[0].Name, resp.Sections[1].Name)
+	}
+	// Redaction-preserved: the redacted flag rides through untouched, and
+	// the failure marker stays a non-redacted marker.
+	if !resp.Sections[0].Redacted {
+		t.Error("config.yaml section should be redacted=true")
+	}
+	if resp.Sections[1].Redacted {
+		t.Error("failure-marker section should be redacted=false")
+	}
+	// Handler does NOT re-derive secrets; stored [REDACTED] verbatim.
+	if want := "ANTHROPIC_API_KEY=[REDACTED]"; !strings.Contains(resp.Sections[0].Content, want) {
+		t.Errorf("section content = %q, want it to contain %q", resp.Sections[0].Content, want)
+	}
+}
+
+// TestGetRescue_404WhenNone — no bundle on file → 404, not 500/200.
+func TestGetRescue_404WhenNone(t *testing.T) {
+	fake := &fakeRescueStore{ret: nil} // store returns (nil, nil)
+	w := doRescueGet(t, "ws-none", "org-9", fake)
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("status = %d, want 404; body=%s", w.Code, w.Body.String())
+	}
+}
+
+// TestGetRescue_OrgScopingPassedToStore — the handler must hand the
+// tenant's MOLECULE_ORG_ID to the store, and a store that returns nil for
+// a mismatched org yields 404. This is the sibling-org isolation: a
+// caller in org B (a different tenant process, MOLECULE_ORG_ID=org-B)
+// reading ws-1 (which belongs to org-9) gets the org filter applied → no
+// row → 404.
+func TestGetRescue_OrgScopingPassedToStore(t *testing.T) {
+	// Tenant configured as a DIFFERENT org than the bundle's owner.
+	// Fake mimics the Postgres org filter: returns nil because org-B
+	// doesn't match the row's org-9.
+	fake := &fakeRescueStore{ret: nil}
+	w := doRescueGet(t, "ws-1", "org-B", fake)
+
+	if fake.gotOrgID != "org-B" {
+		t.Errorf("store got org_id = %q, want the tenant's org-B", fake.gotOrgID)
+	}
+	if fake.gotWorkspaceID != "ws-1" {
+		t.Errorf("store got workspace_id = %q, want ws-1", fake.gotWorkspaceID)
+	}
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("sibling-org read: status = %d, want 404", w.Code)
+	}
+}
+
+// TestGetRescue_EmptyOrgEnvRejected — empty MOLECULE_ORG_ID is a
+// fail-closed security violation (#2020). The handler must 503 before
+// calling the store, so the org filter cannot be bypassed.
+func TestGetRescue_EmptyOrgEnvRejected(t *testing.T) {
+	fake := &fakeRescueStore{ret: sampleStored()}
+	w := doRescueGet(t, "ws-1", "", fake)
+	if fake.gotOrgID != "" {
+		t.Errorf("store was called with org_id = %q; want no call when env empty", fake.gotOrgID)
+	}
+	if w.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status = %d, want 503; body=%s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "platform_misconfigured") {
+		t.Fatalf("body = %s, want platform_misconfigured code", w.Body.String())
+	}
+}
+
+// TestGetRescue_StoreErrorIs503 — an actual datastore error must surface
+// as 503, never a 404 (which would hide an outage as "no bundle").
+func TestGetRescue_StoreErrorIs503(t *testing.T) {
+	fake := &fakeRescueStore{err: errors.New("connection refused")}
+	w := doRescueGet(t, "ws-1", "org-9", fake)
+	if w.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status = %d, want 503; body=%s", w.Code, w.Body.String())
+	}
+}
+
+// TestGetRescue_NilStoreIs503 — defensive: a handler with no store wired
+// (db.DB nil in a degraded boot) returns 503, never panics.
+func TestGetRescue_NilStoreIs503(t *testing.T) {
+	t.Setenv("MOLECULE_ORG_ID", "org-9")
+	h := &RescueReadHandler{} // store == nil
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/rescue", nil)
+	h.GetRescue(c)
+	if w.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status = %d, want 503", w.Code)
+	}
+}
+
+// TestBuildRescueResponse_BoundsSections — a stored bundle with more than
+// maxResponseSections sections is capped + flagged truncated.
+func TestBuildRescueResponse_BoundsSections(t *testing.T) {
+	many := make([]rescue.Section, maxResponseSections+5)
+	for i := range many {
+		many[i] = rescue.Section{Name: "s", Content: "c", Redacted: true}
+	}
+	stored := &rescuestore.StoredBundle{
+		CapturedAt: time.Now(),
+		Bundle:     rescue.Bundle{WorkspaceID: "ws-1", Sections: many},
+	}
+	resp := buildRescueResponse("ws-1", stored)
+	if len(resp.Sections) != maxResponseSections {
+		t.Errorf("sections = %d, want capped at %d", len(resp.Sections), maxResponseSections)
+	}
+	if !resp.Truncated {
+		t.Error("truncated flag should be set when sections were capped")
+	}
+}
@@ -0,0 +1,168 @@
+package handlers
+
+// rescue_wiring.go — bridges the leaf internal/rescue package to the
+// handlers package's EIC/SSH runner + secret redactor, and exposes the
+// boot-failure rescue hook used by both boot-failure verdict paths
+// (handlers.BootstrapFailed here, registry.sweepStuckProvisioning via
+// an injected hook wired in main.go).
+//
+// Why the indirection: internal/rescue is a leaf so registry (which
+// must NOT import handlers — that's an import cycle) can call it. The
+// two heavy dependencies live here in handlers — `withEICTunnel`
+// (the EIC keypair → push → tunnel → ssh dance) and `redactSecrets`
+// (the SAFE-T1201 secret-scan) — so we inject them into rescue's
+// package-level func vars at init().
+//
+// RFC internal#742 Part 2.
+
+import (
+	"bytes"
+	"context"
+	"database/sql"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescuestore"
+)
+
+func init() {
+	// Wire the leaf rescue package to handlers' EIC runner + redactor.
+	// Done in init() (not main.go) so the binding is present for any
+	// caller of rescue.Capture, including the registry sweeper hook and
+	// the handler path, without each call site re-wiring it.
+	rescue.RunRemote = rescueRunRemoteViaEIC
+	rescue.Redact = func(workspaceID, content string) string {
+		out, _ := redactSecrets(workspaceID, content)
+		return out
+	}
+	// Part 3: persist the redacted bundle to the queryable store on
+	// capture so GET /workspaces/:id/rescue can serve it without obs/Loki
+	// read creds. db.DB is resolved per-call (rescuestore guards a nil
+	// handle) so wiring at init() is safe even before InitPostgres has
+	// run; a capture before the DB is up logs + skips the persist rather
+	// than failing the boot-failure path.
+	rescue.PersistBundle = func(ctx context.Context, b rescue.Bundle) error {
+		return rescuestore.NewPostgres(db.DB).Persist(ctx, b)
+	}
+}
+
+// rescueRunRemoteViaEIC runs a single shell command on the still-running
+// (but boot-failed) workspace EC2 over an EIC tunnel and returns its
+// combined stdout+stderr. Reuses the same `withEICTunnel` dance as the
+// canvas file ops (ephemeral keypair → SendSSHPublicKey → open-tunnel →
+// ssh) so the rescue path inherits every fix to the EIC mechanism (e.g.
+// PR #2822's LogLevel=ERROR shim) for free.
+//
+// Combined output (2>&1) is intentional: a boot-failed box's most
+// useful signal is often on stderr (a panic, a missing-file error), and
+// the rescue bundle is a forensic blob, not a parsed value — we want
+// everything the command emitted.
+func rescueRunRemoteViaEIC(ctx context.Context, instanceID, command string) (string, error) {
+	var combined []byte
+	runErr := withEICTunnel(ctx, instanceID, func(s eicSSHSession) error {
+		sshCmd := exec.CommandContext(ctx, "ssh", s.sshArgs(command)...)
+		sshCmd.Env = os.Environ()
+		var buf bytes.Buffer
+		sshCmd.Stdout = &buf
+		sshCmd.Stderr = &buf
+		// A non-zero remote exit is NOT a transport error for the rescue
+		// path — each section command already falls back to an
+		// `|| echo '(...)'` marker, so a clean exit is expected. Only
+		// surface an error when ssh/tunnel itself failed AND produced no
+		// output to ship.
+		err := sshCmd.Run()
+		combined = buf.Bytes()
+		if err != nil && len(combined) == 0 {
+			return fmt.Errorf("rescue ssh exec: %w", err)
+		}
+		return nil
+	})
+	if runErr != nil {
+		return "", runErr
+	}
+	return strings.TrimRight(string(combined), "\n"), nil
+}
+
+// captureRescueBundle fires a best-effort, non-blocking rescue capture
+// for a boot-failed workspace. It is the single entry point both
+// boot-failure verdict paths funnel through.
+//
+// NON-BLOCKING: the actual collection runs in its own goroutine with
+// its own timeout (rescue.CaptureTimeout), detached from the caller's
+// request/sweep context so it can't add latency to — or be cancelled
+// by — the failure-handling path that triggered it. We snapshot the
+// identity into a fresh context.Background() for the same reason: a
+// gin request context is cancelled the instant the HTTP handler
+// returns, which would kill the EIC tunnel mid-collection.
+//
+// instanceID/orgID are resolved here (best-effort) so the two call
+// sites only need the workspace id. A missing instance id → rescue.Capture
+// no-ops (logged), so an early-failure workspace that never got an EC2
+// is handled cleanly.
+func captureRescueBundle(workspaceID, reason string) {
+	rescueDispatch(func() {
+		ctx := context.Background()
+		instanceID, err := rescueResolveInstanceID(ctx, workspaceID)
+		if err != nil {
+			// Best-effort: a resolve failure is logged inside Capture's
+			// caller chain; pass empty so Capture no-ops cleanly.
+			instanceID = ""
+		}
+		rescue.Capture(ctx, rescue.Input{
+			InstanceID:  instanceID,
+			WorkspaceID: workspaceID,
+			OrgID:       os.Getenv("MOLECULE_ORG_ID"),
+			Reason:      reason,
+		})
+	})
+}
+
+// rescueDispatch runs the rescue collection off the request path. In
+// production it's `go fn()` so the capture never blocks or adds latency
+// to the boot-failure handler. Tests swap it for a synchronous runner so
+// they can assert the capture fired (or didn't) deterministically
+// without racing the goroutine.
+var rescueDispatch = func(fn func()) { go fn() }
+
+// BootFailureRescueHook is the registry-facing adapter wired into
+// registry.BootFailureRescueHook from main.go. The registry sweeper
+// already resolved the instance id (it's in the candidate row), so this
+// path uses it directly rather than re-querying — symmetric with the
+// captureRescueBundle handler path but skipping the lookup.
+//
+// Best-effort + non-blocking: dispatches the capture on its own
+// goroutine with its own timeout, so the sweep loop is never slowed.
+func BootFailureRescueHook(workspaceID, instanceID, reason string) {
+	go rescue.Capture(context.Background(), rescue.Input{
+		InstanceID:  instanceID,
+		WorkspaceID: workspaceID,
+		OrgID:       os.Getenv("MOLECULE_ORG_ID"),
+		Reason:      reason,
+	})
+}
+
+// rescueResolveInstanceID looks up the EC2 instance id for a workspace.
+// Package var so tests can stub it without a sqlmock. Mirrors
+// provisioner.resolveInstanceID (same query) but lives here to keep the
+// rescue wiring self-contained and avoid widening the provisioner
+// surface.
+var rescueResolveInstanceID = func(ctx context.Context, workspaceID string) (string, error) {
+	if db.DB == nil {
+		return "", nil // nil in unit tests
+	}
+	var instanceID sql.NullString
+	err := db.DB.QueryRowContext(ctx,
+		`SELECT instance_id FROM workspaces WHERE id = $1`, workspaceID,
+	).Scan(&instanceID)
+	if err != nil && err != sql.ErrNoRows {
+		return "", err
+	}
+	if !instanceID.Valid {
+		return "", nil
+	}
+	return instanceID.String, nil
+}
@@ -0,0 +1,119 @@
+package handlers
+
+import (
+	"bytes"
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/models"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+)
+
+// rescueTestHarness makes the otherwise-async rescue capture
+// deterministic + observable for handler tests:
+//   - rescueDispatch runs synchronously (no goroutine race).
+//   - rescueResolveInstanceID returns a fixed instance id.
+//   - rescue.RunRemote / rescue.Redact are stubbed so no real EIC/SSH
+//     fires; runCalls counts how many remote-command collections ran,
+//     which is the proxy for "did the capture fire".
+//
+// All originals are restored on cleanup.
+func rescueTestHarness(t *testing.T, instanceID string) (runCalls *int) {
+	t.Helper()
+	n := 0
+	runCalls = &n
+
+	prevDispatch := rescueDispatch
+	rescueDispatch = func(fn func()) { fn() } // synchronous
+	prevResolve := rescueResolveInstanceID
+	rescueResolveInstanceID = func(_ context.Context, _ string) (string, error) { return instanceID, nil }
+	prevRun, prevRedact := rescue.RunRemote, rescue.Redact
+	rescue.RunRemote = func(_ context.Context, _ string, _ string) (string, error) { n++; return "out", nil }
+	rescue.Redact = func(_ws, c string) string { return c }
+
+	t.Cleanup(func() {
+		rescueDispatch = prevDispatch
+		rescueResolveInstanceID = prevResolve
+		rescue.RunRemote = prevRun
+		rescue.Redact = prevRedact
+	})
+	return runCalls
+}
+
+// TestBootstrapFailed_FiresRescueOnFlip — the RFC internal#742 handler
+// hook: when BootstrapFailed actually flips a workspace to `failed`
+// (affected==1), the rescue capture fires against the resolved instance.
+func TestBootstrapFailed_FiresRescueOnFlip(t *testing.T) {
+	h, mock := setupBootstrapHandler(t)
+	runCalls := rescueTestHarness(t, "i-failed01")
+
+	mock.ExpectExec(`UPDATE workspaces`).
+		WithArgs("ws-crashed", sqlmock.AnyArg(), models.StatusFailed).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+	mock.ExpectExec(`INSERT INTO structure_events`).
+		WithArgs("WORKSPACE_PROVISION_FAILED", "ws-crashed", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-crashed"}}
+	c.Request = httptest.NewRequest("POST", "/admin/workspaces/ws-crashed/bootstrap-failed",
+		bytes.NewBufferString(`{"error":"codex provider derivation failed","log_tail":"panic"}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	h.BootstrapFailed(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("want 200, got %d: %s", w.Code, w.Body.String())
+	}
+	if *runCalls != len(rescueBundleSectionCount()) {
+		t.Errorf("rescue capture ran %d remote commands, want %d (one per bundle section)", *runCalls, len(rescueBundleSectionCount()))
+	}
+}
+
+// TestBootstrapFailed_NoRescueOnNoChange — an already-transitioned
+// workspace (affected==0: raced to online, or double-report) is NOT a
+// boot-failure verdict here, so the rescue capture must NOT fire.
+func TestBootstrapFailed_NoRescueOnNoChange(t *testing.T) {
+	h, mock := setupBootstrapHandler(t)
+	runCalls := rescueTestHarness(t, "i-online01")
+
+	mock.ExpectExec(`UPDATE workspaces`).
+		WithArgs("ws-online", sqlmock.AnyArg(), models.StatusFailed).
+		WillReturnResult(sqlmock.NewResult(0, 0)) // already transitioned
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-online"}}
+	c.Request = httptest.NewRequest("POST", "/admin/workspaces/ws-online/bootstrap-failed",
+		bytes.NewBufferString(`{"error":"late report","log_tail":""}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	h.BootstrapFailed(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("want 200, got %d", w.Code)
+	}
+	if *runCalls != 0 {
+		t.Errorf("rescue capture fired (%d cmds) on a no-change report; it must only fire on a real flip", *runCalls)
+	}
+}
+
+// rescueBundleSectionCount returns the production rescue bundle section
+// list length by running a capture against a counting runner once. It's
+// a small indirection so the handler test stays decoupled from the exact
+// section set in internal/rescue (which has its own tests).
+func rescueBundleSectionCount() []struct{} {
+	count := 0
+	prevRun, prevRedact := rescue.RunRemote, rescue.Redact
+	rescue.RunRemote = func(_ context.Context, _ string, _ string) (string, error) { count++; return "", nil }
+	rescue.Redact = func(_ws, c string) string { return c }
+	rescue.Capture(context.Background(), rescue.Input{InstanceID: "i-probe", WorkspaceID: "w", OrgID: "o"})
+	rescue.RunRemote = prevRun
+	rescue.Redact = prevRedact
+	return make([]struct{}, count)
+}
@@ -176,6 +176,10 @@ func TestResolveAgentURLForRestartSignal_CacheMiss(t *testing.T) {
 // TestGracefulPreRestart_Success verifies that when the workspace returns 200,
 // the signal is logged as acknowledged without error.
 func TestGracefulPreRestart_Success(t *testing.T) {
+	hWrapper := &resolveURLTestWrapper{
+		WorkspaceHandler: newHandlerWithTestDeps(t),
+		testURL:          "http://fake-agent.example/agent",
+	}
 	_ = setupTestDB(t)

 	// httptest server simulating the workspace container's /signals/restart_pending
@@ -205,18 +209,15 @@ func TestGracefulPreRestart_Success(t *testing.T) {
 		})
 	}))
 	defer srv.Close()
+	hWrapper.testURL = srv.URL + "/agent"

 	// Pre-populate Redis cache with the test server URL
 	_ = setupTestRedisWithURL(t, srv.URL)

-	// Use a wrapper so gracefulPreRestart runs through the embedded handler.
-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		testURL:          srv.URL + "/agent",
-	}
+	// gracefulPreRestart runs in a goroutine; wait for it before db.DB is restored.
+	// Must be registered AFTER setupTestDB (LIFO: async wait → db.DB restore).
+	waitForHandlerAsyncBeforeDBCleanup(t, hWrapper.WorkspaceHandler)

-	// gracefulPreRestart runs in a goroutine with its own timeout.
-	// We give it time to complete before the test ends.
 	hWrapper.gracefulPreRestart(context.Background(), "ws-ack-789")
 	time.Sleep(200 * time.Millisecond)
 }
@@ -224,19 +225,22 @@ func TestGracefulPreRestart_Success(t *testing.T) {
 // TestGracefulPreRestart_NotImplemented verifies that when the workspace returns
 // 404 (old SDK version), the platform proceeds gracefully (log + no error).
 func TestGracefulPreRestart_NotImplemented(t *testing.T) {
+	hWrapper := &resolveURLTestWrapper{
+		WorkspaceHandler: newHandlerWithTestDeps(t),
+		testURL:          "http://fake-agent.example/agent",
+	}
 	_ = setupTestDB(t)

 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusNotFound)
 	}))
 	defer srv.Close()
+	hWrapper.testURL = srv.URL + "/agent"

 	_ = setupTestRedisWithURL(t, srv.URL)

-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		testURL:          srv.URL + "/agent",
-	}
+	// Must be registered AFTER setupTestDB so LIFO order is: async wait → db.DB restore.
+	waitForHandlerAsyncBeforeDBCleanup(t, hWrapper.WorkspaceHandler)

 	hWrapper.gracefulPreRestart(context.Background(), "ws-noimpl-999")
 	time.Sleep(200 * time.Millisecond)
@@ -246,15 +250,18 @@ func TestGracefulPreRestart_NotImplemented(t *testing.T) {
 // TestGracefulPreRestart_ConnectionRefused verifies that when the workspace
 // is unreachable, the platform proceeds gracefully without error.
 func TestGracefulPreRestart_ConnectionRefused(t *testing.T) {
-	_ = setupTestDB(t)
-
-	mr := setupTestRedisWithURL(t, "http://localhost:19999/agent") // nothing listening on 19999
-	_ = mr
-
 	hWrapper := &resolveURLTestWrapper{
 		WorkspaceHandler: newHandlerWithTestDeps(t),
 		testURL:          "http://localhost:19999/agent",
 	}
+	_ = setupTestDB(t)
+
+	// Nothing listening on 19999 — deliberate connection failure.
+	mr := setupTestRedisWithURL(t, "http://localhost:19999/agent")
+	_ = mr
+
+	// Must be registered AFTER setupTestDB so LIFO order is: async wait → db.DB restore.
+	waitForHandlerAsyncBeforeDBCleanup(t, hWrapper.WorkspaceHandler)

 	hWrapper.gracefulPreRestart(context.Background(), "ws-unreachable-000")
 	time.Sleep(200 * time.Millisecond)
@@ -264,13 +271,17 @@ func TestGracefulPreRestart_ConnectionRefused(t *testing.T) {
 // TestGracefulPreRestart_URLResolutionError verifies that when URL resolution
 // fails, the platform proceeds gracefully without blocking the restart.
 func TestGracefulPreRestart_URLResolutionError(t *testing.T) {
-	_ = setupTestDB(t)
-	_ = setupTestRedis(t) // empty → URL resolution will fail in resolveAgentURLForRestartSignal
-
 	hWrapper := &resolveURLTestWrapper{
 		WorkspaceHandler: newHandlerWithTestDeps(t),
 		errToReturn:      context.DeadlineExceeded,
 	}
+	_ = setupTestDB(t)
+	_ = setupTestRedis(t) // empty → URL resolution will fail in resolveAgentURLForRestartSignal
+
+	// Must be registered AFTER setupTestDB so LIFO order is: async wait → db.DB restore.
+	// This ensures goroutines (which access both DB and Redis) are drained before
+	// any cleanup fires. setupTestRedis comes after newHandlerWithTestDeps
+	// so the handler holds the correct Redis client reference.
 	waitForHandlerAsyncBeforeDBCleanup(t, hWrapper.WorkspaceHandler)

 	hWrapper.gracefulPreRestart(context.Background(), "ws-url-err-111")
@@ -44,6 +44,20 @@ func billingModeForRegistryProvider(p providers.Provider) string {
 	return LLMBillingModeBYOK
 }

+// requiredEnvForRegistryProvider derives the env vars the USER must supply for
+// a model owned by the resolved provider — the proper-SSOT single fact behind
+// the canvas "Missing API Keys" preflight (task #65). The closed platform
+// provider injects credentials server-side (the metered proxy) -> nothing
+// required; BYOK providers require their auth_env. Derived from IsPlatform +
+// AuthEnv so a template can no longer hand-author a required_env that drifts
+// from the registry's serving classification.
+func requiredEnvForRegistryProvider(p providers.Provider) []string {
+	if p.IsPlatform() {
+		return nil
+	}
+	return p.AuthEnv
+}
+
 // enrichFromRegistry populates the registry-served fields on a templateSummary
 // when its runtime is known to the provider registry. It is a no-op (leaves
 // RegistryBacked=false and the registry slices nil) for a runtime the registry
@@ -98,6 +112,7 @@ func enrichFromRegistry(summary *templateSummary, runtime string) {
 		if derived, derr := m.DeriveProvider(runtime, id, nil); derr == nil {
 			ms.Provider = derived.Name
 			ms.BillingMode = billingModeForRegistryProvider(derived)
+			ms.RequiredEnv = requiredEnvForRegistryProvider(derived)
 		}
 		// If DeriveProvider errors (ambiguous/overlap — a manifest defect the
 		// loader's tests pin against), still serve the id without a provider
@@ -1,12 +1,15 @@
 package handlers

 import (
+	"context"
 	"fmt"
 	"io"
 	"net/http"
 	"os"
 	"time"

+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/crypto"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
 	"github.com/gin-gonic/gin"
 )

@@ -18,30 +21,69 @@ func NewTracesHandler() *TracesHandler {
 	return &TracesHandler{}
 }

+type langfuseConfig struct {
+	Host   string
+	Public string
+	Secret string
+}
+
+// resolveLangfuseConfig resolves Langfuse connection settings from
+// admin-controlled global secrets first, then process env for legacy/dev use.
+// Workspace secrets are intentionally excluded: a workspace-controlled
+// LANGFUSE_HOST would allow SSRF with BasicAuth attached (#2029).
+func resolveLangfuseConfig(ctx context.Context) (*langfuseConfig, error) {
+	cfg := &langfuseConfig{}
+
+	resolve := func(key string) string {
+		var val []byte
+		var ver int
+		err := db.DB.QueryRowContext(ctx,
+			`SELECT encrypted_value, encryption_version FROM global_secrets WHERE key = $1`,
+			key).Scan(&val, &ver)
+		if err == nil {
+			decrypted, decErr := crypto.DecryptVersioned(val, ver)
+			if decErr == nil {
+				return string(decrypted)
+			}
+		}
+		return os.Getenv(key)
+	}
+
+	cfg.Host = resolve("LANGFUSE_HOST")
+	cfg.Public = resolve("LANGFUSE_PUBLIC_KEY")
+	cfg.Secret = resolve("LANGFUSE_SECRET_KEY")
+
+	if cfg.Host == "" || cfg.Public == "" || cfg.Secret == "" {
+		return nil, nil
+	}
+	return cfg, nil
+}
+
 // List handles GET /workspaces/:id/traces
 // Proxies to Langfuse API to get recent traces for a workspace.
 func (h *TracesHandler) List(c *gin.Context) {
 	workspaceID := c.Param("id")

-	langfuseHost := os.Getenv("LANGFUSE_HOST")
-	langfusePublic := os.Getenv("LANGFUSE_PUBLIC_KEY")
-	langfuseSecret := os.Getenv("LANGFUSE_SECRET_KEY")
-
-	if langfuseHost == "" || langfusePublic == "" || langfuseSecret == "" {
+	cfg, err := resolveLangfuseConfig(c.Request.Context())
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to resolve trace config"})
+		return
+	}
+	if cfg == nil {
 		c.JSON(http.StatusOK, []interface{}{})
 		return
 	}

 	// Fetch traces from Langfuse, filtered by workspace tag or name
 	url := fmt.Sprintf("%s/api/public/traces?limit=20&orderBy=timestamp&orderDir=desc&tags=%s",
-		langfuseHost, workspaceID)
+		cfg.Host, workspaceID)

 	req, err := http.NewRequestWithContext(c.Request.Context(), "GET", url, nil)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to create request"})
 		return
 	}
-	req.SetBasicAuth(langfusePublic, langfuseSecret)
+	req.SetBasicAuth(cfg.Public, cfg.Secret)

 	resp, err := langfuseClient.Do(req)
 	if err != nil {
@@ -51,10 +93,14 @@ func (h *TracesHandler) List(c *gin.Context) {
 	}
 	defer func() { _ = resp.Body.Close() }()

-	body, err := io.ReadAll(resp.Body)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to read response body"})
+	body, readErr := io.ReadAll(resp.Body)
+	if readErr != nil {
+		c.JSON(http.StatusOK, []interface{}{})
 		return
 	}
-	c.Data(resp.StatusCode, "application/json", body)
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		c.JSON(http.StatusOK, []interface{}{})
+		return
+	}
+	c.Data(http.StatusOK, "application/json", body)
 }
@@ -1,27 +1,31 @@
 package handlers

 import (
+	"database/sql"
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
 	"testing"

+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/crypto"
+	"github.com/DATA-DOG/go-sqlmock"
 	"github.com/gin-gonic/gin"
 )

 // ==================== GET /workspaces/:id/traces ====================

 func TestTracesList_NoLangfuseConfig(t *testing.T) {
-	setupTestDB(t)
+	mock := setupTestDB(t)
 	setupTestRedis(t)
 	handler := NewTracesHandler()

-	// Ensure Langfuse env vars are not set
 	os.Unsetenv("LANGFUSE_HOST")
 	os.Unsetenv("LANGFUSE_PUBLIC_KEY")
 	os.Unsetenv("LANGFUSE_SECRET_KEY")

+	expectMissingGlobalLangfuseSecrets(mock)
+
 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
 	c.Params = gin.Params{{Key: "id", Value: "ws-traces"}}
@@ -41,14 +45,16 @@ func TestTracesList_NoLangfuseConfig(t *testing.T) {
 	if len(resp) != 0 {
 		t.Errorf("expected empty list when Langfuse not configured, got %d items", len(resp))
 	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
 }

 func TestTracesList_PartialLangfuseConfig(t *testing.T) {
-	setupTestDB(t)
+	mock := setupTestDB(t)
 	setupTestRedis(t)
 	handler := NewTracesHandler()

-	// Set only host, missing keys
 	os.Setenv("LANGFUSE_HOST", "http://localhost:3000")
 	os.Unsetenv("LANGFUSE_PUBLIC_KEY")
 	os.Unsetenv("LANGFUSE_SECRET_KEY")
@@ -56,6 +62,8 @@ func TestTracesList_PartialLangfuseConfig(t *testing.T) {
 		os.Unsetenv("LANGFUSE_HOST")
 	}()

+	expectMissingGlobalLangfuseSecrets(mock)
+
 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
 	c.Params = gin.Params{{Key: "id", Value: "ws-traces-partial"}}
@@ -72,10 +80,13 @@ func TestTracesList_PartialLangfuseConfig(t *testing.T) {
 	if len(resp) != 0 {
 		t.Errorf("expected empty list with partial config, got %d items", len(resp))
 	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
 }

 func TestTracesList_LangfuseUnreachable(t *testing.T) {
-	setupTestDB(t)
+	mock := setupTestDB(t)
 	setupTestRedis(t)
 	handler := NewTracesHandler()

@@ -89,6 +100,8 @@ func TestTracesList_LangfuseUnreachable(t *testing.T) {
 		os.Unsetenv("LANGFUSE_SECRET_KEY")
 	}()

+	expectMissingGlobalLangfuseSecrets(mock)
+
 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
 	c.Params = gin.Params{{Key: "id", Value: "ws-traces-down"}}
@@ -106,4 +119,171 @@ func TestTracesList_LangfuseUnreachable(t *testing.T) {
 	if len(resp) != 0 {
 		t.Errorf("expected empty list when Langfuse unreachable, got %d items", len(resp))
 	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestTracesList_GlobalSecretsFallback(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewTracesHandler()
+
+	os.Unsetenv("LANGFUSE_HOST")
+	os.Unsetenv("LANGFUSE_PUBLIC_KEY")
+	os.Unsetenv("LANGFUSE_SECRET_KEY")
+
+	expectGlobalLangfuseSecret(mock, "LANGFUSE_HOST", "http://localhost:3000")
+	expectGlobalLangfuseSecret(mock, "LANGFUSE_PUBLIC_KEY", "pk-global")
+	expectGlobalLangfuseSecret(mock, "LANGFUSE_SECRET_KEY", "sk-global")
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-traces-global"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-traces-global/traces", nil)
+
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var resp []interface{}
+	json.Unmarshal(w.Body.Bytes(), &resp)
+	if len(resp) != 0 {
+		t.Errorf("expected empty list when Langfuse unreachable, got %d items", len(resp))
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestTracesList_GlobalPartialConfig(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewTracesHandler()
+
+	os.Unsetenv("LANGFUSE_HOST")
+	os.Unsetenv("LANGFUSE_PUBLIC_KEY")
+	os.Unsetenv("LANGFUSE_SECRET_KEY")
+
+	expectGlobalLangfuseSecret(mock, "LANGFUSE_HOST", "http://localhost:3000")
+	mock.ExpectQuery(`SELECT encrypted_value, encryption_version FROM global_secrets WHERE key = \$1`).
+		WithArgs("LANGFUSE_PUBLIC_KEY").
+		WillReturnError(sql.ErrNoRows)
+	mock.ExpectQuery(`SELECT encrypted_value, encryption_version FROM global_secrets WHERE key = \$1`).
+		WithArgs("LANGFUSE_SECRET_KEY").
+		WillReturnError(sql.ErrNoRows)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-traces-partial"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-traces-partial/traces", nil)
+
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var resp []interface{}
+	json.Unmarshal(w.Body.Bytes(), &resp)
+	if len(resp) != 0 {
+		t.Errorf("expected empty list with partial config, got %d items", len(resp))
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestTracesList_LangfuseUpstreamError(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewTracesHandler()
+
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		w.WriteHeader(http.StatusInternalServerError)
+		w.Write([]byte("<html><body>Internal Server Error</body></html>"))
+	}))
+	defer upstream.Close()
+
+	os.Setenv("LANGFUSE_HOST", upstream.URL)
+	os.Setenv("LANGFUSE_PUBLIC_KEY", "pk-test")
+	os.Setenv("LANGFUSE_SECRET_KEY", "sk-test")
+	defer func() {
+		os.Unsetenv("LANGFUSE_HOST")
+		os.Unsetenv("LANGFUSE_PUBLIC_KEY")
+		os.Unsetenv("LANGFUSE_SECRET_KEY")
+	}()
+
+	expectMissingGlobalLangfuseSecrets(mock)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-traces-500"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-traces-500/traces", nil)
+
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var resp []interface{}
+	json.Unmarshal(w.Body.Bytes(), &resp)
+	if len(resp) != 0 {
+		t.Errorf("expected empty list on upstream error, got %d items", len(resp))
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func TestTracesList_WorkspaceSecretsIgnored(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewTracesHandler()
+
+	os.Unsetenv("LANGFUSE_HOST")
+	os.Unsetenv("LANGFUSE_PUBLIC_KEY")
+	os.Unsetenv("LANGFUSE_SECRET_KEY")
+
+	expectMissingGlobalLangfuseSecrets(mock)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-traces-ssrf"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-traces-ssrf/traces", nil)
+
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var resp []interface{}
+	json.Unmarshal(w.Body.Bytes(), &resp)
+	if len(resp) != 0 {
+		t.Errorf("expected empty list when workspace secrets ignored, got %d items", len(resp))
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations not met: %v", err)
+	}
+}
+
+func expectMissingGlobalLangfuseSecrets(mock sqlmock.Sqlmock) {
+	for _, key := range []string{"LANGFUSE_HOST", "LANGFUSE_PUBLIC_KEY", "LANGFUSE_SECRET_KEY"} {
+		mock.ExpectQuery(`SELECT encrypted_value, encryption_version FROM global_secrets WHERE key = \$1`).
+			WithArgs(key).
+			WillReturnError(sql.ErrNoRows)
+	}
+}
+
+func expectGlobalLangfuseSecret(mock sqlmock.Sqlmock, key, value string) {
+	enc, _ := crypto.Encrypt([]byte(value))
+	ver := crypto.CurrentEncryptionVersion()
+	mock.ExpectQuery(`SELECT encrypted_value, encryption_version FROM global_secrets WHERE key = \$1`).
+		WithArgs(key).
+		WillReturnRows(sqlmock.NewRows([]string{"encrypted_value", "encryption_version"}).AddRow(enc, ver))
 }
@@ -474,6 +474,32 @@ func (h *WorkspaceHandler) Create(c *gin.Context) {
 			})
 			return
 		}
+		// issue #2172 (provider-side companion): once the (runtime, model)
+		// pair is known to be registered, confirm the DERIVED provider
+		// (the one the adapter will resolve at boot) is a known provider
+		// in the providers.yaml catalog. Live trigger (adk-demo Assistant,
+		// 2026-06-03): the model-side check passed for a registry-resident
+		// model whose derived provider name was NOT in the providers list,
+		// so the save was accepted and the agent wedged at boot with
+		// "provider=X not in providers registry". This check is a
+		// defense-in-depth registry-consistency guard: by construction a
+		// model in a runtime's native set derives to a provider that IS in
+		// the catalog, so the rejection path is primarily a registry-data
+		// invariant — any future drift between `providers:` and `runtimes:`
+		// fails the create with a clear pointer to the missing provider
+		// rather than silently wedging the agent. Fails open for runtimes
+		// the registry doesn't know (langgraph/external/kimi/mock/federated
+		// — the federation contract the model-side check also honors).
+		if ok, why := validateDerivedProviderInRegistry(payload.Runtime, payload.Model); !ok {
+			log.Printf("Create: 422 DERIVED_PROVIDER_NOT_IN_REGISTRY (runtime=%q model=%q): %s [issue #2172 hard-reject]", payload.Runtime, payload.Model, why)
+			c.JSON(http.StatusUnprocessableEntity, gin.H{
+				"error":   why,
+				"runtime": payload.Runtime,
+				"model":   payload.Model,
+				"code":    "DERIVED_PROVIDER_NOT_IN_REGISTRY",
+			})
+			return
+		}
 	}

 	ctx := c.Request.Context()
@@ -56,7 +56,20 @@ func PatchAbilities(c *gin.Context) {
 		return
 	}

-	if body.BroadcastEnabled != nil {
+	// Atomic update: when both fields are supplied, apply them in one SQL
+	// statement so the request is all-or-nothing (#2131). A partial mutation
+	// (e.g. broadcast_enabled updated but talk_to_user_enabled failing) would
+	// leave the workspace in an ambiguous capability state.
+	if body.BroadcastEnabled != nil && body.TalkToUserEnabled != nil {
+		if _, err := db.DB.ExecContext(ctx,
+			`UPDATE workspaces SET broadcast_enabled = $2, talk_to_user_enabled = $3, updated_at = now() WHERE id = $1`,
+			id, *body.BroadcastEnabled, *body.TalkToUserEnabled,
+		); err != nil {
+			log.Printf("PatchAbilities both-fields for %s: %v", id, err)
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "update failed"})
+			return
+		}
+	} else if body.BroadcastEnabled != nil {
 		if _, err := db.DB.ExecContext(ctx,
 			`UPDATE workspaces SET broadcast_enabled = $2, updated_at = now() WHERE id = $1`,
 			id, *body.BroadcastEnabled,
@@ -65,9 +78,7 @@ func PatchAbilities(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "update failed"})
 			return
 		}
-	}
-
-	if body.TalkToUserEnabled != nil {
+	} else if body.TalkToUserEnabled != nil {
 		if _, err := db.DB.ExecContext(ctx,
 			`UPDATE workspaces SET talk_to_user_enabled = $2, updated_at = now() WHERE id = $1`,
 			id, *body.TalkToUserEnabled,
@@ -130,11 +130,8 @@ func TestPatchAbilities_BothFields(t *testing.T) {
 	mock.ExpectQuery(`SELECT EXISTS\(SELECT 1 FROM workspaces WHERE id = \$1 AND status != 'removed'\)`).
 		WithArgs(wsUUID1).
 		WillReturnRows(sqlmock.NewRows([]string{"exists"}).AddRow(true))
-	mock.ExpectExec(`UPDATE workspaces SET broadcast_enabled = \$2, updated_at = now\(\) WHERE id = \$1`).
-		WithArgs(wsUUID1, true).
-		WillReturnResult(sqlmock.NewResult(0, 1))
-	mock.ExpectExec(`UPDATE workspaces SET talk_to_user_enabled = \$2, updated_at = now\(\) WHERE id = \$1`).
-		WithArgs(wsUUID1, true).
+	mock.ExpectExec(`UPDATE workspaces SET broadcast_enabled = \$2, talk_to_user_enabled = \$3, updated_at = now\(\) WHERE id = \$1`).
+		WithArgs(wsUUID1, true, true).
 		WillReturnResult(sqlmock.NewResult(0, 1))

 	w := patchAbilitiesReq(t, wsUUID1, `{"broadcast_enabled":true,"talk_to_user_enabled":true}`)
@@ -182,19 +179,25 @@ func TestPatchAbilities_TalkToUserUpdateError(t *testing.T) {
 	}
 }

-func TestPatchAbilities_BothFields_BroadcastFails(t *testing.T) {
+// TestPatchAbilities_BothFields_UpdateError — regression for #2131. When
+// both fields are supplied the handler uses a single combined UPDATE. A
+// failure of that UPDATE must leave the workspace unchanged (atomic).
+func TestPatchAbilities_BothFields_UpdateError(t *testing.T) {
 	mock, cleanup := withMockDB(t)
 	defer cleanup()

 	mock.ExpectQuery(`SELECT EXISTS\(SELECT 1 FROM workspaces WHERE id = \$1 AND status != 'removed'\)`).
 		WithArgs(wsUUID1).
 		WillReturnRows(sqlmock.NewRows([]string{"exists"}).AddRow(true))
-	mock.ExpectExec(`UPDATE workspaces SET broadcast_enabled = \$2, updated_at = now\(\) WHERE id = \$1`).
-		WithArgs(wsUUID1, true).
+	mock.ExpectExec(`UPDATE workspaces SET broadcast_enabled = \$2, talk_to_user_enabled = \$3, updated_at = now\(\) WHERE id = \$1`).
+		WithArgs(wsUUID1, true, true).
 		WillReturnError(errors.New("disk full"))

 	w := patchAbilitiesReq(t, wsUUID1, `{"broadcast_enabled":true,"talk_to_user_enabled":true}`)
 	if w.Code != http.StatusInternalServerError {
 		t.Fatalf("expected 500, got %d: %s", w.Code, w.Body.String())
 	}
+	// Because only one UPDATE is issued, there is no partial-mutation
+	// path to assert against; sqlmock implicitly verifies no second
+	// exec occurred.
 }
@@ -91,6 +91,18 @@ func (h *WorkspaceHandler) BootstrapFailed(c *gin.Context) {
 		"log_tail": tail,
 		"source":   "bootstrap_watcher",
 	})
+
+	// RFC internal#742 Part 2: this is one of the two boot-failure
+	// verdict points. We've just flipped a still-running (but
+	// unconfigured) workspace EC2 to `failed`; the control plane will
+	// reap the instance shortly. Capture a forensic rescue bundle off
+	// the live box NOW, before it's torn down, so a wedged workspace is
+	// post-mortem-inspectable. Best-effort + non-blocking: runs in its
+	// own goroutine with its own timeout, detached from this request's
+	// context (which is cancelled the instant this handler returns).
+	// Failure to capture never changes the boot-failure handling.
+	captureRescueBundle(id, "bootstrap_watcher")
+
 	log.Printf("BootstrapFailed: marked %s failed (tail=%d bytes, err=%q)", id, len(tail), errMsg)
 	c.JSON(http.StatusOK, gin.H{"ok": true})
 }
@@ -95,6 +95,14 @@ func TestIntegration_BroadcastOrgRoot_NonRootSenderResolvesToRoot(t *testing.T)
 		}
 	})

+	// Pre-test hygiene: if a prior run crashed or was killed, its rows may
+	// still be in the shared integration DB. Remove them before inserting so
+	// the unique index workspaces_parent_name_uniq does not conflict.
+	if _, err := conn.ExecContext(ctx,
+		`DELETE FROM workspaces WHERE name LIKE $1`, prefix+"%"); err != nil {
+		t.Logf("pre-test cleanup (non-fatal): %v", err)
+	}
+
 	rootID := uuid.New().String()
 	midID := uuid.New().String()
 	leafID := uuid.New().String()
@@ -41,7 +41,6 @@ import (
 	"context"
 	"database/sql"
 	"fmt"
-	"os"
 	"testing"

 	"github.com/google/uuid"
@@ -59,10 +58,7 @@ import (
 // only those.
 func integrationDB_WorkspaceCreateName(t *testing.T) *sql.DB {
 	t.Helper()
-	url := os.Getenv("INTEGRATION_DB_URL")
-	if url == "" {
-		t.Skip("INTEGRATION_DB_URL not set; skipping (see file header)")
-	}
+	url := requireIntegrationDBURL(t)
 	conn, err := sql.Open("postgres", url)
 	if err != nil {
 		t.Fatalf("open: %v", err)
@@ -93,3 +93,16 @@ func formatMissingBYOKCredentialError(mode string) string {
 		mode,
 	)
 }
+
+// formatMissingPlatformProxyError builds the user-facing message for a
+// provision failure caused by a platform-managed workspace whose control-plane
+// proxy environment is absent (#2162). The platform-managed path requires
+// MOLECULE_LLM_BASE_URL + MOLECULE_LLM_USAGE_TOKEN (or their OPENAI_*
+// fallbacks) to inject a usable credential; without them the workspace must
+// NOT start credential-less.
+func formatMissingPlatformProxyError() string {
+	return "this workspace is configured for platform-managed LLM billing but the control-plane proxy is not ready. " +
+		"The required platform proxy env (MOLECULE_LLM_BASE_URL + MOLECULE_LLM_USAGE_TOKEN) is absent. " +
+		"This is usually a transient boot-race; retry in 30 seconds. If it persists, verify the platform proxy " +
+		"is configured for this tenant/runtime and contact the platform team."
+}
@@ -613,6 +613,32 @@ func (h *WorkspaceHandler) ensureDefaultConfig(workspaceID string, payload model
 	if model == "" {
 		log.Printf("ensureDefaultConfig: workspace %s reached provisioning with empty model — Create handler should have rejected this; rendering empty model: \"\" in config.yaml (workspace will boot not_configured)", workspaceID)
 	}
+
+	// Derive the provider from the providers manifest and stamp it into the
+	// generated config BEFORE claude-code model normalization strips the
+	// slash-prefix. DeriveProvider needs the FULL, un-normalized model id
+	// (e.g. "moonshot/kimi-k2.6") for the exact-id match that resolves the
+	// canvas claude-code case to provider=platform — normalizing to
+	// "kimi-k2.6" first would lose that match.
+	//
+	// Why this exists (RFC#340 Fix A): a canvas-created claude-code workspace
+	// with model "moonshot/kimi-k2.6" booted NOT_CONFIGURED — the adapter
+	// derived provider="moonshot" (slash-split of the model id) which is not
+	// in the providers registry. CP bakes `provider: platform` via heredoc,
+	// but the cp#329 config-bundle fetch overwrites /configs/config.yaml with
+	// THIS (previously providerless) bundle version, so molecule-runtime
+	// config.py re-derived the wrong provider. Stamping the manifest-derived
+	// provider here (mirroring CP's buildModelProviderYAML shape) makes the
+	// config the adapter reads carry the canonical provider.
+	//
+	// Reuses the SAME manifest path the config-SAVE validators use
+	// (providerRegistry() + Manifest.DeriveProvider; see
+	// model_registry_validation.go). On a derive MISS (unknown/unregistered
+	// model, or registry unavailable) provider is left empty and the field is
+	// omitted below — preserving today's behavior; never fail provisioning on
+	// a derive miss.
+	derivedProvider := deriveDefaultConfigProvider(runtime, model)
+
 	if runtime == "claude-code" {
 		model = normalizeClaudeCodeModel(model)
 	}
@@ -640,6 +666,14 @@ func (h *WorkspaceHandler) ensureDefaultConfig(workspaceID string, payload model
 	// Model always at top level — config.py reads raw["model"] for all runtimes.
 	configYAML += fmt.Sprintf("model: %s\n", quoteModel)

+	// Stamp the manifest-derived provider at top level (mirroring CP's
+	// buildModelProviderYAML). Omitted entirely on a derive miss so the prior
+	// behavior — no `provider:` key, runtime re-derives — is preserved for
+	// unregistered models (requirement #3).
+	if derivedProvider != "" {
+		configYAML += fmt.Sprintf("provider: '%s'\n", yamlEscapeSingleQuotedProvider(derivedProvider))
+	}
+
 	// Add runtime_config. required_env is intentionally omitted — the
 	// platform injects secrets at container-start time via the secrets API,
 	// and preflight already validates that the env vars are present before
@@ -649,6 +683,10 @@ func (h *WorkspaceHandler) ensureDefaultConfig(workspaceID string, payload model
 	if runtime == "claude-code" {
 		configYAML += fmt.Sprintf("  model: %s\n", quoteModel)
 	}
+	// Mirror the top-level provider under runtime_config (CP writes both).
+	if derivedProvider != "" {
+		configYAML += fmt.Sprintf("  provider: '%s'\n", yamlEscapeSingleQuotedProvider(derivedProvider))
+	}
 	configYAML += "  timeout: 0\n"

 	files["config.yaml"] = []byte(configYAML)
@@ -657,6 +695,48 @@ func (h *WorkspaceHandler) ensureDefaultConfig(workspaceID string, payload model
 	return files
 }

+// deriveDefaultConfigProvider resolves the provider name the adapter should
+// see for (runtime, model) using the SAME providers manifest the config-SAVE
+// validators use (providerRegistry() + Manifest.DeriveProvider; see
+// model_registry_validation.go). It is intentionally fail-OPEN: any miss
+// (empty model, registry unavailable, unknown runtime, or a model the runtime
+// does not own) returns "" so the caller omits the `provider:` field and the
+// generated config keeps its pre-fix shape. It NEVER fails provisioning.
+//
+// `model` must be the FULL, un-normalized id (e.g. "moonshot/kimi-k2.6") so
+// DeriveProvider's exact-id match resolves the canvas claude-code case to
+// provider=platform. The availableAuthEnv arg is nil here — config-generation
+// has no per-workspace auth context yet (secrets are injected at container
+// start), matching the validators' nil call.
+func deriveDefaultConfigProvider(runtime, model string) string {
+	if strings.TrimSpace(model) == "" {
+		return ""
+	}
+	m, err := providerRegistry()
+	if err != nil || m == nil {
+		// Registry unavailable (a build-time defect the gen/sync gates catch).
+		// Fail open — do not stamp a provider, do not block provisioning.
+		return ""
+	}
+	p, err := m.DeriveProvider(runtime, model, nil)
+	if err != nil {
+		// Unknown runtime (federation / non-first-party) or a model the
+		// runtime does not own. Either way, omit the provider and let the
+		// runtime fall back to its prior derivation — preserving today's
+		// behavior for unregistered models.
+		return ""
+	}
+	return p.Name
+}
+
+// yamlEscapeSingleQuotedProvider escapes a value for a YAML single-quoted
+// scalar, mirroring CP's buildModelProviderYAML (a literal single quote is
+// doubled). Provider names are registry-controlled identifiers, so this is a
+// defense-in-depth measure rather than a hot path.
+func yamlEscapeSingleQuotedProvider(v string) string {
+	return strings.ReplaceAll(v, "'", "''")
+}
+
 func normalizeClaudeCodeModel(model string) string {
 	model = strings.TrimSpace(model)
 	if before, after, ok := strings.Cut(model, "/"); ok && before != "" && after != "" {
@@ -1003,12 +1083,13 @@ func applyPlatformManagedLLMEnv(ctx context.Context, envVars map[string]string,
 	anthropicBaseURL := firstNonEmptyEnv("MOLECULE_LLM_ANTHROPIC_BASE_URL", "ANTHROPIC_BASE_URL")
 	token := firstNonEmptyEnv("MOLECULE_LLM_USAGE_TOKEN", "OPENAI_API_KEY")
 	if baseURL == "" || token == "" {
-		// Proxy not configured (boot race / misconfig). On the platform_managed
-		// path the workspace IS entitled to platform creds, so we do NOT strip
-		// here — but we report HasUsableLLMCred from whatever survived so the
-		// caller's fail-closed branch (non-platform only) is never reached on
-		// this path.
-		return platformLLMEnvResult{ResolvedMode: res.ResolvedMode, HasUsableLLMCred: true, Source: res.Source}
+		// Proxy not configured (boot race / misconfig). The platform_managed
+		// path REQUIRES the CP proxy env to inject a usable credential.
+		// Reporting HasUsableLLMCred=true here would start the workspace
+		// credential-less — the adk-demo dark-wedge class (#2162).
+		// Return false so the caller's fail-closed branch aborts with
+		// MISSING_PLATFORM_PROXY.
+		return platformLLMEnvResult{ResolvedMode: res.ResolvedMode, HasUsableLLMCred: false, Source: res.Source}
 	}
 	stripPlatformManagedLLMBypassEnv(envVars)

@@ -134,6 +134,11 @@ func TestProvisionWorkspaceAuto_NoBackendMarksFailed(t *testing.T) {
 // This is the regression-prevention test for the Design Director bug
 // where 7-of-7 sub-agents went down the Docker path on SaaS.
 func TestProvisionWorkspaceAuto_RoutesToCPWhenSet(t *testing.T) {
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	mock := setupTestDB(t)
 	mock.MatchExpectationsInOrder(false)

@@ -597,6 +602,11 @@ func TestNoCallSiteCallsBareStop(t *testing.T) {
 // count without mocking out the retry helper itself, which would
 // invert the test contract — the retry IS the dispatcher's job here).
 func TestRestartWorkspaceAuto_RoutesToCPWhenSet(t *testing.T) {
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	rec := &trackingCPProv{}
 	bcast := &concurrentSafeBroadcaster{}
 	h := NewWorkspaceHandler(bcast, nil, "http://localhost:8080", t.TempDir())
@@ -795,6 +805,11 @@ func TestResumeHandler_UsesProvisionWorkspaceAuto(t *testing.T) {
 // the async tests; the absence of `go` semantics is the load-bearing
 // distinction we're pinning.
 func TestProvisionWorkspaceAutoSync_RoutesToCPWhenSet(t *testing.T) {
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	mock := setupTestDB(t)
 	mock.MatchExpectationsInOrder(false)
 	// provisionWorkspaceCP runs prepareProvisionContext synchronously, which
@@ -98,6 +98,11 @@ func (r *recordingCPProv) startedSet() map[string]struct{} {
 func TestProvisionWorkspaceCP_ConcurrentBurst_NoSilentDrop(t *testing.T) {
 	const numWorkspaces = 7

+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	mock := setupTestDB(t)

 	// Every goroutine runs prepareProvisionContext → mintWorkspaceSecrets
@@ -0,0 +1,196 @@
+package handlers
+
+// workspace_provision_platform_boot_test.go — the deterministic, SSOT-driven
+// regression suite for the class of bug behind the moonshot/kimi
+// "canvas-created claude-code workspace boots NOT_CONFIGURED" production
+// incident (RFC#340 Fix A #2187, canvas Fix C #2188).
+//
+// THE BUG (what shipped to prod):
+//   A claude-code workspace created via the canvas with provider=Platform +
+//   model="moonshot/kimi-k2.6" booted NOT_CONFIGURED. Unit tests passed; the
+//   REAL boot path was broken. ensureDefaultConfig generated a config.yaml that
+//   carried NO derived `provider:` key, so the cp#329 config-bundle the adapter
+//   actually reads left molecule-runtime config.py to slash-split the model id
+//   "moonshot/kimi-k2.6" -> provider="moonshot", which is NOT in the providers
+//   registry -> NOT_CONFIGURED.
+//
+// THE FIX A INVARIANT (this file pins it, and pins it for the WHOLE class):
+//   ensureDefaultConfig MUST stamp the manifest-derived provider into the
+//   generated config.yaml — at BOTH the top level and under runtime_config —
+//   for every (runtime, model) the providers SSOT maps to a platform provider.
+//   The single-combo pin (TestEnsureDefaultConfig_StampsDerivedProvider in
+//   workspace_provision_test.go) proves the headline case. THIS file closes the
+//   gap that single pin leaves: it is PARAMETRIZED OVER THE SSOT, so when a NEW
+//   platform model is added to providers.yaml for claude-code (or any runtime
+//   with a platform arm), the new id is automatically covered — a future
+//   platform model that fails to derive `provider: platform` fails THIS test at
+//   build time, before it can ship a NOT_CONFIGURED boot.
+//
+// WHY SSOT-DRIVEN AND NOT A HAND-MAINTAINED LIST:
+//   The original bug was a divergence between "what the canvas offers"
+//   (providers.yaml platform arm) and "what the config generator stamps". A
+//   hardcoded test model list would itself drift from the SSOT and re-open the
+//   same divergence gap. By enumerating the platform model set directly from the
+//   loaded providers.Manifest (the SAME manifest ensureDefaultConfig's
+//   deriveDefaultConfigProvider resolves against), this test cannot fall behind
+//   the offered set: add a platform model, get a test case for free; the test
+//   only passes if the generator actually stamps it.
+//
+// SCOPE: deterministic, no live infra. The REAL-boot complement (provision a
+// staging workspace and assert status=online + a completion returns 200 for the
+// SAME combo) is the bash staging harness — see
+// tests/e2e/test_staging_full_saas.sh (E2E_LLM_PATH=platform) and the
+// e2e-staging-platform-boot job in .gitea/workflows/e2e-staging-saas.yml. That
+// asserts the REAL artifact (booted status / completion); THIS asserts the
+// deterministic config-generation invariant the real boot depends on.
+
+import (
+	"testing"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/models"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/providers"
+	"gopkg.in/yaml.v3"
+)
+
+// platformModelsForRuntime returns the exact model ids the providers SSOT lists
+// under runtime rt's `platform` native provider arm — the set the canvas offers
+// as provider=Platform and the set ensureDefaultConfig MUST stamp
+// `provider: platform` for. Reads the SAME embedded manifest the config
+// generator derives against (providers.LoadManifest), so it can never drift from
+// the offered set. Returns nil when the runtime has no platform arm.
+func platformModelsForRuntime(t *testing.T, rt string) []string {
+	t.Helper()
+	m, err := providers.LoadManifest()
+	if err != nil {
+		t.Fatalf("LoadManifest: %v", err)
+	}
+	native, ok := m.Runtimes[rt]
+	if !ok {
+		t.Fatalf("providers SSOT has no runtimes entry for %q", rt)
+	}
+	for _, ref := range native.Providers {
+		if ref.Name == "platform" {
+			return ref.Models
+		}
+	}
+	return nil
+}
+
+// TestEnsureDefaultConfig_StampsProviderForEverySSOTPlatformModel is the
+// class-level regression for the moonshot/kimi NOT_CONFIGURED incident. For
+// EVERY model the providers SSOT offers under claude-code's platform arm, it
+// asserts the generated config.yaml carries the manifest-derived provider at
+// both the top level and under runtime_config. This is the Fix A invariant,
+// parametrized over the SSOT so a newly-offered platform model cannot ship
+// without the stamp (the exact divergence — offered-but-not-stamped — that
+// booted "moonshot/kimi-k2.6" into NOT_CONFIGURED).
+func TestEnsureDefaultConfig_StampsProviderForEverySSOTPlatformModel(t *testing.T) {
+	const runtime = "claude-code"
+	platformModels := platformModelsForRuntime(t, runtime)
+	if len(platformModels) == 0 {
+		t.Fatalf("providers SSOT lists no platform models for runtime %q — the regression matrix would be empty; the SSOT shape changed (this test is the canary)", runtime)
+	}
+	// Headline sentinel: the exact id that booted NOT_CONFIGURED in prod MUST be
+	// in the enumerated set. If a refactor drops it from the platform arm, this
+	// test must still cover it explicitly — fail loud rather than silently
+	// shrinking the matrix.
+	if !containsString(platformModels, "moonshot/kimi-k2.6") {
+		t.Fatalf("the headline incident model \"moonshot/kimi-k2.6\" is no longer in the claude-code platform SSOT set (%v) — regression coverage for the original bug would be lost", platformModels)
+	}
+
+	for _, model := range platformModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			broadcaster := newTestBroadcaster()
+			handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
+
+			files := handler.ensureDefaultConfig("ws-platform-boot", models.CreateWorkspacePayload{
+				Name:    "Platform Boot Agent",
+				Tier:    2,
+				Runtime: runtime,
+				Model:   model,
+			})
+
+			raw, ok := files["config.yaml"]
+			if !ok {
+				t.Fatalf("expected config.yaml in generated files for model %q", model)
+			}
+
+			var parsed struct {
+				Model         string `yaml:"model"`
+				Provider      string `yaml:"provider"`
+				RuntimeConfig struct {
+					Model    string `yaml:"model"`
+					Provider string `yaml:"provider"`
+				} `yaml:"runtime_config"`
+			}
+			if err := yaml.Unmarshal(raw, &parsed); err != nil {
+				t.Fatalf("generated YAML invalid for model %q: %v\n%s", model, err, raw)
+			}
+
+			// The load-bearing invariant: BOTH the top-level and the
+			// runtime_config provider must be exactly "platform". An empty or
+			// vendor-namespace ("moonshot") value here is the prod NOT_CONFIGURED
+			// boot — the adapter would slash-split the model id and look up an
+			// unregistered provider.
+			if parsed.Provider != "platform" {
+				t.Errorf("model %q: top-level provider = %q, want \"platform\" (Fix A invariant — empty/vendor value is the NOT_CONFIGURED boot)\n%s", model, parsed.Provider, raw)
+			}
+			if parsed.RuntimeConfig.Provider != "platform" {
+				t.Errorf("model %q: runtime_config.provider = %q, want \"platform\"\n%s", model, parsed.RuntimeConfig.Provider, raw)
+			}
+			// Sanity: the config must still render a non-empty model (a config
+			// with provider but no model is equally undeployable).
+			if parsed.Model == "" {
+				t.Errorf("model %q: generated config has empty top-level model\n%s", model, raw)
+			}
+		})
+	}
+}
+
+// TestPlatformModelDeriveProvider_SSOTConsistency is the upstream half of the
+// same invariant, one layer below ensureDefaultConfig: it asserts the providers
+// manifest's DeriveProvider — the resolver deriveDefaultConfigProvider calls —
+// maps every SSOT-offered claude-code platform model to a provider whose Name is
+// "platform". If DeriveProvider itself regressed (e.g. a model_prefix_match
+// change made "moonshot/kimi-k2.6" resolve to the bare "moonshot" entry again),
+// this fails closer to the root cause than the config-shape test above, making
+// the diagnosis unambiguous: SSOT/derive regression vs config-emission
+// regression.
+func TestPlatformModelDeriveProvider_SSOTConsistency(t *testing.T) {
+	const runtime = "claude-code"
+	m, err := providers.LoadManifest()
+	if err != nil {
+		t.Fatalf("LoadManifest: %v", err)
+	}
+	platformModels := platformModelsForRuntime(t, runtime)
+	if len(platformModels) == 0 {
+		t.Fatalf("no platform models for %q in SSOT", runtime)
+	}
+	for _, model := range platformModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			// nil availableAuthEnv mirrors deriveDefaultConfigProvider's call at
+			// config-generation time (no per-workspace auth context yet).
+			p, err := m.DeriveProvider(runtime, model, nil)
+			if err != nil {
+				t.Fatalf("DeriveProvider(%q, %q): unexpected error %v — an SSOT-offered platform model MUST derive", runtime, model, err)
+			}
+			if p.Name != "platform" {
+				t.Errorf("DeriveProvider(%q, %q).Name = %q, want \"platform\" (this is the exact slash-split-to-vendor regression that booted NOT_CONFIGURED)", runtime, model, p.Name)
+			}
+		})
+	}
+}
+
+// containsString is a tiny local membership helper. Kept here (not a shared
+// test util) so this regression file is self-contained and can be read top to
+// bottom without chasing helpers across the package.
+func containsString(xs []string, want string) bool {
+	for _, x := range xs {
+		if x == want {
+			return true
+		}
+	}
+	return false
+}
@@ -230,6 +230,18 @@ func (h *WorkspaceHandler) prepareProvisionContext(
 			Extra: map[string]interface{}{"error": msg, "code": "MISSING_BYOK_CREDENTIAL", "billing_mode": llmRes.ResolvedMode, "issue": "1994"},
 		}
 	}
+	// Fail closed for a platform-managed workspace whose CP proxy env is
+	// absent: do NOT start it credential-less (adk-demo dark-wedge class,
+	// #2162). The platform_managed path requires the proxy injection to
+	// produce a usable credential.
+	if llmRes.ResolvedMode == LLMBillingModePlatformManaged && !llmRes.HasUsableLLMCred {
+		msg := formatMissingPlatformProxyError()
+		log.Printf("Provisioner: ABORT workspace=%s — platform-managed billing mode but CP proxy env absent (MISSING_PLATFORM_PROXY, molecule-core#2162)", workspaceID)
+		return nil, &provisionAbort{
+			Msg:   msg,
+			Extra: map[string]interface{}{"error": msg, "code": "MISSING_PLATFORM_PROXY", "billing_mode": llmRes.ResolvedMode, "issue": "2162"},
+		}
+	}
 	applyRuntimeModelEnv(envVars, payload.Runtime, payload.Model)
 	if payload.Role != "" {
 		envVars["MOLECULE_AGENT_ROLE"] = payload.Role
@@ -264,6 +264,11 @@ func TestPrepareProvisionContext_ParentIDInjection(t *testing.T) {
 		},
 	}

+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
 			mock := setupTestDB(t)
@@ -331,6 +336,10 @@ func TestPrepareProvisionContext_InjectsGitHTTPCredsFromPersonaToken(t *testing.
 		}
 	}
 	t.Setenv("MOLECULE_PERSONA_ROOT", root)
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")

 	cases := []struct {
 		name         string
@@ -459,6 +468,10 @@ func TestPrepareProvisionContext_WorkspaceSecretWinsOverPersonaToken(t *testing.
 		t.Fatal(err)
 	}
 	t.Setenv("MOLECULE_PERSONA_ROOT", root)
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")

 	mock := setupTestDB(t)
 	mock.ExpectQuery(`SELECT key, encrypted_value, encryption_version FROM global_secrets`).
@@ -363,6 +363,74 @@ runtime_config:
 	}
 }

+// TestEnsureDefaultConfig_StampsDerivedProvider pins RFC#340 Fix A: a
+// canvas-created claude-code workspace with model "moonshot/kimi-k2.6" must
+// have the manifest-derived provider stamped into config.yaml at BOTH the top
+// level and under runtime_config, so the cp#329 config-bundle the adapter
+// reads no longer leaves the runtime to slash-split "moonshot/..." → an
+// unregistered provider="moonshot" (the original NOT_CONFIGURED boot). The
+// canonical manifest exact-id-matches "moonshot/kimi-k2.6" to provider=platform.
+func TestEnsureDefaultConfig_StampsDerivedProvider(t *testing.T) {
+	broadcaster := newTestBroadcaster()
+	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
+
+	files := handler.ensureDefaultConfig("ws-moonshot", models.CreateWorkspacePayload{
+		Name:    "Kimi Agent",
+		Tier:    2,
+		Runtime: "claude-code",
+		Model:   "moonshot/kimi-k2.6",
+	})
+
+	var parsed struct {
+		Model         string `yaml:"model"`
+		Provider      string `yaml:"provider"`
+		RuntimeConfig struct {
+			Model    string `yaml:"model"`
+			Provider string `yaml:"provider"`
+		} `yaml:"runtime_config"`
+	}
+	if err := yaml.Unmarshal(files["config.yaml"], &parsed); err != nil {
+		t.Fatalf("generated YAML invalid: %v\n%s", err, files["config.yaml"])
+	}
+	if parsed.Provider != "platform" {
+		t.Errorf("top-level provider = %q, want platform\n%s", parsed.Provider, files["config.yaml"])
+	}
+	if parsed.RuntimeConfig.Provider != "platform" {
+		t.Errorf("runtime_config.provider = %q, want platform\n%s", parsed.RuntimeConfig.Provider, files["config.yaml"])
+	}
+	// The claude-code model normalization still strips the slash prefix.
+	if parsed.Model != "kimi-k2.6" {
+		t.Errorf("top-level model = %q, want kimi-k2.6\n%s", parsed.Model, files["config.yaml"])
+	}
+}
+
+// TestEnsureDefaultConfig_DeriveMissOmitsProvider pins requirement #3: a model
+// the providers manifest does NOT recognize for the runtime (a derive miss)
+// must NOT write any `provider:` key — neither top-level nor under
+// runtime_config — preserving the pre-fix behavior (no empty `provider:`,
+// provisioning never fails on a miss). "gpt-4o" is not a registered
+// claude-code model, so DeriveProvider errors and the field is omitted.
+func TestEnsureDefaultConfig_DeriveMissOmitsProvider(t *testing.T) {
+	broadcaster := newTestBroadcaster()
+	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
+
+	files := handler.ensureDefaultConfig("ws-derivemiss", models.CreateWorkspacePayload{
+		Name:    "Unregistered Agent",
+		Tier:    1,
+		Runtime: "claude-code",
+		Model:   "gpt-4o",
+	})
+
+	content := string(files["config.yaml"])
+	if strings.Contains(content, "provider:") {
+		t.Errorf("derive miss must NOT write any provider: key, got:\n%s", content)
+	}
+	// Sanity: a derive miss must still produce a valid, model-bearing config.
+	if !strings.Contains(content, `model: "gpt-4o"`) {
+		t.Errorf("derive miss should still render the model, got:\n%s", content)
+	}
+}
+
 func TestEnsureDefaultConfig_CustomModel(t *testing.T) {
 	broadcaster := newTestBroadcaster()
 	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
@@ -1424,6 +1492,11 @@ func (s *stubFailingCPProv) IsRunning(_ context.Context, _ string) (bool, error)
 // the broadcast payload would surface every marker; the canned
 // "provisioning failed" message must surface none of them.
 func TestProvisionWorkspaceCP_NoInternalErrorsInBroadcast(t *testing.T) {
+	// Supply the CP proxy env so the platform-managed default does not abort
+	// with MISSING_PLATFORM_PROXY (molecule-core#2162).
+	t.Setenv("MOLECULE_LLM_BASE_URL", "https://api.example.test/api/v1/internal/llm/openai/v1")
+	t.Setenv("MOLECULE_LLM_USAGE_TOKEN", "tenant-admin-token")
+
 	mock := setupTestDB(t)

 	// loadWorkspaceSecrets queries global_secrets and workspace_secrets
@@ -876,8 +876,9 @@ func (h *WorkspaceHandler) runRestartCycle(workspaceID string) {
 	h.provisionWorkspaceAutoSync(workspaceID, "", nil, payload)
 	// sendRestartContext is a one-way notification to the new container; safe
 	// to fire async — the next restart cycle won't depend on it completing.
-	// Tracked via goAsync so the test harness can drain it before the
-	// global db.DB swap (sendRestartContext reads db.DB).
+	// Tracked via h.goAsync so tests can wait for it via h.asyncWG before
+	// closing the sqlmock. Without this, untracked goroutines hit the restored
+	// mock and cause "was not expected" errors in parallel CI execution (mc#1264).
 	h.goAsync(func() { h.sendRestartContext(workspaceID, restartData) })
 }

@@ -16,7 +16,7 @@ const SchemaVersion = 1
 // Fingerprint is a stable content hash of the generated projection (schema
 // version + provider catalog + runtime native sets). It changes iff the
 // registry DATA changes (comment-only YAML edits do not churn it).
-const Fingerprint = "8f733b112695b926"
+const Fingerprint = "a491f5ff8a17ef59"

 // GenProvider is the generated projection of one provider catalog entry —
 // the subset a downstream consumer needs to derive + display a provider.
@@ -50,7 +50,7 @@ var Providers = []GenProvider{
 	{Name: "openai-api", DisplayName: "OpenAI API", Protocol: "openai", AuthMode: "anthropic_api", AuthEnv: []string{"OPENAI_API_KEY"}, ModelPrefixMatch: "^openai-api[:/]", IsPlatform: false, UpstreamVendor: "openai"},
 	{Name: "moonshot", DisplayName: "Moonshot (Kimi)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MOONSHOT_API_KEY", "KIMI_API_KEY"}, ModelPrefixMatch: "^moonshot[:/-]", IsPlatform: false, UpstreamVendor: "moonshot"},
 	{Name: "minimax", DisplayName: "MiniMax", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MINIMAX_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^minimax-m", IsPlatform: false, UpstreamVendor: "minimax"},
-	{Name: "platform", DisplayName: "Platform", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ANTHROPIC_API_KEY", "MOLECULE_LLM_USAGE_TOKEN"}, ModelPrefixMatch: "^platform/", IsPlatform: true},
+	{Name: "platform", DisplayName: "Platform", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MOLECULE_LLM_USAGE_TOKEN"}, ModelPrefixMatch: "^platform/", IsPlatform: true},
 	{Name: "xiaomi-mimo", DisplayName: "Xiaomi MiMo", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "^mimo-", IsPlatform: false},
 	{Name: "zai", DisplayName: "Z.ai (GLM)", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GLM_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^glm-", IsPlatform: false},
 	{Name: "kimi-coding", DisplayName: "Moonshot Kimi (coding-tuned)", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"KIMI_API_KEY", "ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN"}, ModelPrefixMatch: "^kimi-", IsPlatform: false},
@@ -89,8 +89,9 @@ var Runtimes = map[string][]GenRuntimeRef{
 		{Name: "platform", Models: []string{"openai/gpt-5.4", "openai/gpt-5.4-mini"}},
 	},
 	"google-adk": {
+		{Name: "platform", Models: []string{"platform:gemini-2.5-pro", "platform:gemini-2.5-flash"}},
+		{Name: "google", Models: []string{"gemini-2.5-pro", "gemini-2.5-flash"}},
 		{Name: "vertex", Models: []string{"vertex:gemini-2.5-pro"}},
-		{Name: "google", Models: []string{"gemini-2.5-pro"}},
 	},
 	"hermes": {
 		{Name: "kimi-coding", Models: []string{"kimi-coding/kimi-k2"}},
@@ -0,0 +1,42 @@
+package providers
+
+import "testing"
+
+// Proper-SSOT (task #65): google-adk keyless Gemini resolves to the closed
+// platform provider -> IsPlatform=true; BYOK AI Studio -> google. The
+// platform: select ids are registered so workspace-create accepts them
+// (was 422 UNREGISTERED_MODEL_FOR_RUNTIME).
+func TestGoogleADK_PlatformGeminiResolvesToPlatform(t *testing.T) {
+	m, err := LoadManifest()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, id := range []string{"platform:gemini-2.5-pro", "platform:gemini-2.5-flash"} {
+		p, err := m.DeriveProvider("google-adk", id, nil)
+		if err != nil {
+			t.Fatalf("%s: %v", id, err)
+		}
+		if p.Name != PlatformProviderName || !p.IsPlatform() {
+			t.Errorf("%s -> %q IsPlatform=%v; want platform", id, p.Name, p.IsPlatform())
+		}
+	}
+	p, err := m.DeriveProvider("google-adk", "gemini-2.5-pro", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if p.IsPlatform() || p.Name != "google" {
+		t.Errorf("gemini-2.5-pro -> %q IsPlatform=%v; want google byok", p.Name, p.IsPlatform())
+	}
+	models, _ := m.ModelsForRuntime("google-adk")
+	want := map[string]bool{"platform:gemini-2.5-pro": false, "platform:gemini-2.5-flash": false}
+	for _, id := range models {
+		if _, ok := want[id]; ok {
+			want[id] = true
+		}
+	}
+	for id, ok := range want {
+		if !ok {
+			t.Errorf("%s not registered for google-adk — create would 422", id)
+		}
+	}
+}
@@ -292,7 +292,7 @@ providers:
    # PR-1 simplification when only claude-code referenced platform.
    base_url_template: "https://api.moleculesai.app/api/v1/internal/llm/openai/v1"
    base_url_anthropic: "https://api.moleculesai.app/api/v1/internal/llm/anthropic/v1"
-    auth_env: [ANTHROPIC_API_KEY, MOLECULE_LLM_USAGE_TOKEN]
+    auth_env: [MOLECULE_LLM_USAGE_TOKEN]
    auth_token_env: ANTHROPIC_API_KEY
    # Adapter routes kimi- / moonshot/ through platform by default. No bare
    # vendor prefix of its own; it multiplexes other vendors' slugs. Match
@@ -412,14 +412,24 @@ providers:
    model_prefix_match: "^gemini-"
    model_aliases: []

-  # Google Vertex AI — KEYLESS arm (mirrors the anthropic-oauth / anthropic-api
-  # and openai-subscription / openai-api split: same vendor, distinct auth).
-  # google-adk serves Gemini via Vertex using Application Default Credentials
-  # over Workload Identity Federation (AWS EC2 role -> GCP STS -> SA), injected
-  # by the provisioner (cp#416 + envs.yaml vertex block) as a NON-SECRET
-  # external_account cred-config at GOOGLE_APPLICATION_CREDENTIALS. No API key.
-  # Distinct `vertex:` model namespace keeps it unambiguous vs the API-key
-  # `google` vendor's ^gemini- (TestNoAmbiguousModelMatch).
+  # Google Vertex AI — served via the Molecule CP LLM proxy (NOT on-box ADC).
+  # google-adk routes ALL Gemini through the proxy, which mints a Vertex token
+  # server-side over Workload Identity Federation (AWS -> GCP STS -> SA; see
+  # internal/vertexauth + llm_proxy.go google/vertex case) and meters usage to
+  # org credits. The former on-box ADC delivery (a NON-SECRET external_account
+  # cred-config written to /configs/gcp-adc.json via GOOGLE_APPLICATION_CREDENTIALS,
+  # injected by the provisioner) was REMOVED: a tenant has EC2 root and could
+  # have used that credential to call Vertex DIRECTLY, bypassing metering
+  # (keyless-Vertex billing leak — task #64 / RFC internal#763; provisioner
+  # force-off + template routes vertex: through the proxy). The `vertex:`
+  # namespace + this entry remain for proxy routing/billing of Vertex-upstream
+  # requests, distinct from the API-key `google` vendor's ^gemini-
+  # (TestNoAmbiguousModelMatch).
+  #
+  # NOTE: display_name ("keyless ADC") and auth_env (GOOGLE_APPLICATION_CREDENTIALS)
+  # are now VESTIGIAL — no consumer reads auth_env post-leak-fix, but it must stay
+  # non-empty (providers.go validate). Left as-is to keep this a comment-only,
+  # regen-free change; retiring them is a registry-regen follow-up.
  - name: vertex
    display_name: "Google Vertex AI (keyless ADC)"
    vendor_logo: "google"
@@ -844,11 +854,24 @@ runtimes:
  # this runtimes entry declares the selectable model set.
  google-adk:
    providers:
-      # Keyless Vertex (org-compliant default): Gemini via Vertex AI + ADC/WIF.
-      - name: vertex
+      # Platform-managed (keyless, metered) Gemini via the Molecule LLM proxy ->
+      # Vertex AI (server-side WIF mint; NO credential on the tenant box — the
+      # keyless-Vertex leak fix). Resolves to the closed `platform` provider ->
+      # IsPlatform=true -> platform_managed billing + required_env=[] (derived).
+      # The org-compliant default. The runtime translates the `platform:` select
+      # id to the bare wire id the proxy routes to Vertex.
+      - name: platform
        models:
-          - vertex:gemini-2.5-pro
-      # API-key BYOK arm: AI Studio GEMINI_API_KEY/GOOGLE_API_KEY.
+          - platform:gemini-2.5-pro
+          - platform:gemini-2.5-flash
+      # API-key BYOK arm: AI Studio (the tenant's OWN GOOGLE_API_KEY).
      - name: google
        models:
-          - gemini-2.5-pro
+          - gemini-2.5-pro
+          - gemini-2.5-flash
+      # DEPRECATED transitional: vertex: ids stay registered until templates
+      # move to platform: (superseded by the platform arm above). Remove in a
+      # cleanup once no template references vertex:gemini-*.
+      - name: vertex
+        models:
+          - vertex:gemini-2.5-pro
@@ -29,7 +29,7 @@ import (
 // canonicalProvidersYAMLSHA256 is the sha256 of the canonical providers.yaml as
 // synced from molecule-controlplane. Bumped deliberately on each re-sync (see
 // file doc). Cross-checked live by the sync-providers-yaml CI workflow.
-const canonicalProvidersYAMLSHA256 = "dec73199e26cee2d395a0acece99771618d3879dc5ca724ba57cb5b38079c6ce"
+const canonicalProvidersYAMLSHA256 = "021ae082c2bbbbb61c406cae03205ac6b7fff160ae5976cfc64de3de676d02b2"

 func TestSyncedYAMLMatchesCanonicalSHA(t *testing.T) {
 	sum := sha256.Sum256(embeddedYAML)
@@ -10,8 +10,15 @@ import (
 	"github.com/DATA-DOG/go-sqlmock"

 	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
 )

+// rescueVolumeGraceHours surfaces the rescue grace as whole hours for
+// the retention-contract assertion (RFC internal#742 Part 2).
+func rescueVolumeGraceHours() int {
+	return int(rescue.RescueVolumeGrace.Hours())
+}
+
 // fakeCPReaper is a hand-rolled CPOrphanReaper for the SaaS-mode
 // sweeper tests. Records every Stop call so tests can assert which
 // workspace IDs were re-issued.
@@ -97,6 +104,55 @@ func TestCPSweepOnce_NoOrphans(t *testing.T) {
 	}
 }

+// TestCPSweepOnce_DoesNotReapFailedWorkspace — RFC internal#742 Part 2
+// volume-retention guarantee, molecule-core side.
+//
+// A boot-FAILED workspace (status='failed') must NOT be terminated by
+// the platform's orphan sweeper: its instance + /configs data volume are
+// retained through the rescue grace (rescue.RescueVolumeGrace) so a live
+// rescue read is possible, distinct from the user-prune erase path. The
+// sweeper reaps ONLY status='removed' (the explicit deprovision path),
+// so a `failed` row is structurally excluded at the SELECT — it never
+// reaches reaper.Stop. We assert the predicate filters to 'removed'
+// (so the failed instance survives) and that no Stop fires for a DB
+// whose only orphan-shaped row is `failed`.
+//
+// This is the "if the sweeper already keeps volumes by default, confirm
+// + add a test asserting it" branch of the RFC: it does, by construction.
+func TestCPSweepOnce_DoesNotReapFailedWorkspace(t *testing.T) {
+	mock := setupTestDB(t)
+	reaper := &fakeCPReaper{}
+
+	// The sweeper's SELECT carries `status = 'removed'`. A boot-failed
+	// workspace (status='failed') does not match that predicate, so the
+	// real DB returns it nowhere in this result set — modelled as the
+	// empty result the `removed`-only filter produces when the only
+	// instance-bearing row is `failed`. The regex pins the retention-
+	// critical predicate so a future widening to include 'failed' (which
+	// would terminate boot-failed boxes mid-rescue) fails this test.
+	mock.ExpectQuery(`(?s)WHERE status = 'removed'\s+AND instance_id IS NOT NULL`).
+		WithArgs(cpSweepLimit).
+		WillReturnRows(sqlmock.NewRows([]string{"id"})) // failed row excluded by predicate
+
+	cpSweepOnce(context.Background(), reaper)
+
+	if len(reaper.stopCalls) != 0 {
+		t.Fatalf("boot-failed workspace must be RETAINED (no terminate); got Stop calls %v", reaper.stopCalls)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestRescueVolumeGraceIsDistinctFromPrune documents that the rescue
+// grace is its own contract (24h) and not coupled to any prune timing —
+// the value is the SSOT the control-plane reaper must honour.
+func TestRescueVolumeGraceIsDistinctFromPrune(t *testing.T) {
+	if rescueVolumeGraceHours() != 24 {
+		t.Errorf("rescue volume grace = %dh, want 24h (RFC internal#742)", rescueVolumeGraceHours())
+	}
+}
+
 // TestCPSweepOnce_MultipleOrphans — all rows in the batch get Stop'd
 // independently; one failure doesn't block others.
 func TestCPSweepOnce_MultipleOrphans(t *testing.T) {
@@ -92,6 +92,23 @@ func provisioningTimeoutFor(runtime string, lookup RuntimeTimeoutLookup) time.Du
 	return DefaultProvisioningTimeout
 }

+// BootFailureRescueHook, when wired, is invoked once per workspace the
+// sweep flips from `provisioning` to `failed` — i.e. on the boot-failure
+// verdict, BEFORE the control plane reaps the instance. It captures a
+// forensic rescue bundle off the still-running (but boot-failed) EC2 and
+// ships it to obs/Loki (RFC internal#742 Part 2). Wired in main.go to
+// handlers.captureRescueBundle via a thin adapter; nil in tests + on
+// self-hosted deploys (no rescue shipping there).
+//
+// Function-typed injection (not an import of handlers) keeps the
+// existing handlers→registry import direction intact — registry must not
+// import handlers.
+//
+// MUST be best-effort + non-blocking: the hook itself dispatches the
+// capture on its own goroutine with its own timeout, so the sweep loop
+// is never slowed or blocked by a hung EIC tunnel on the dead box.
+var BootFailureRescueHook func(workspaceID, instanceID, reason string)
+
 // StartProvisioningTimeoutSweep periodically scans for workspaces stuck in
 // `status='provisioning'` past the timeout window, flips them to `failed`,
 // and broadcasts a WORKSPACE_PROVISION_TIMEOUT event so the canvas can
@@ -144,7 +161,7 @@ func sweepStuckProvisioning(ctx context.Context, emitter ProvisionTimeoutEmitter
 	// flight, not historical) and the partial index on status keeps
 	// it fast.
 	rows, err := db.DB.QueryContext(ctx, `
-		SELECT id, COALESCE(runtime, ''), EXTRACT(EPOCH FROM (now() - updated_at))::int
+		SELECT id, COALESCE(runtime, ''), COALESCE(instance_id, ''), EXTRACT(EPOCH FROM (now() - updated_at))::int
 		FROM workspaces
 		WHERE status = 'provisioning'
 	`)
@@ -155,14 +172,15 @@ func sweepStuckProvisioning(ctx context.Context, emitter ProvisionTimeoutEmitter
 	defer rows.Close()

 	type candidate struct {
-		id      string
-		runtime string
-		ageSec  int
+		id         string
+		runtime    string
+		instanceID string
+		ageSec     int
 	}
 	var ids []candidate
 	for rows.Next() {
 		var c candidate
-		if err := rows.Scan(&c.id, &c.runtime, &c.ageSec); err == nil {
+		if err := rows.Scan(&c.id, &c.runtime, &c.instanceID, &c.ageSec); err == nil {
 			ids = append(ids, c)
 		}
 	}
@@ -200,6 +218,19 @@ func sweepStuckProvisioning(ctx context.Context, emitter ProvisionTimeoutEmitter
 			continue
 		}
 		log.Printf("Provision-timeout sweep: %s (runtime=%q) stuck in provisioning > %s — marked failed", c.id, c.runtime, timeout)
+
+		// RFC internal#742 Part 2: this flip is a boot-failure verdict.
+		// The instance is still running (the CP reaps it shortly after);
+		// capture a forensic rescue bundle off it NOW, before teardown.
+		// Best-effort + non-blocking — the hook dispatches on its own
+		// goroutine + timeout, so a hung EIC tunnel on the dead box can't
+		// slow the sweep. Only fires on a real flip (affected==1), never
+		// on a race (affected==0) or a non-overdue row — guaranteeing it
+		// runs once per boot-failure verdict and never on a healthy row.
+		if BootFailureRescueHook != nil {
+			BootFailureRescueHook(c.id, c.instanceID, "provision_timeout_sweep")
+		}
+
 		// Emit as WORKSPACE_PROVISION_FAILED, not _TIMEOUT, because the
 		// canvas event handler only flips node state on the _FAILED case.
 		// A separate event type was considered but the UI reaction is
@@ -0,0 +1,130 @@
+package registry
+
+import (
+	"context"
+	"sync"
+	"testing"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/models"
+	"github.com/DATA-DOG/go-sqlmock"
+)
+
+// rescueHookRecorder captures the args of every BootFailureRescueHook
+// invocation so tests can assert the rescue capture fires exactly on the
+// boot-failure verdict — and never on a healthy/raced row.
+type rescueHookRecorder struct {
+	mu    sync.Mutex
+	calls [][3]string // {workspaceID, instanceID, reason}
+}
+
+func (r *rescueHookRecorder) hook() func(workspaceID, instanceID, reason string) {
+	return func(workspaceID, instanceID, reason string) {
+		r.mu.Lock()
+		defer r.mu.Unlock()
+		r.calls = append(r.calls, [3]string{workspaceID, instanceID, reason})
+	}
+}
+
+func (r *rescueHookRecorder) count() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return len(r.calls)
+}
+
+// withRescueHook installs a recorder as the package-level
+// BootFailureRescueHook for the test's duration.
+func withRescueHook(t *testing.T) *rescueHookRecorder {
+	t.Helper()
+	rec := &rescueHookRecorder{}
+	prev := BootFailureRescueHook
+	BootFailureRescueHook = rec.hook()
+	t.Cleanup(func() { BootFailureRescueHook = prev })
+	return rec
+}
+
+// TestSweep_RescueFiresOnBootFailureVerdict — the core RFC internal#742
+// assertion: when the sweep flips a stuck workspace to `failed`, the
+// rescue hook fires once with the workspace + instance id and the
+// provision_timeout_sweep reason, BEFORE teardown.
+func TestSweep_RescueFiresOnBootFailureVerdict(t *testing.T) {
+	mock := setupTestDB(t)
+	rec := withRescueHook(t)
+
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-stuck", "codex", "i-0badf00d", 700}))
+	mock.ExpectExec(`UPDATE workspaces`).
+		WithArgs("ws-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	sweepStuckProvisioning(context.Background(), &fakeEmitter{}, nil)
+
+	if rec.count() != 1 {
+		t.Fatalf("rescue hook should fire once on a boot-failure flip, got %d", rec.count())
+	}
+	got := rec.calls[0]
+	if got[0] != "ws-stuck" || got[1] != "i-0badf00d" || got[2] != "provision_timeout_sweep" {
+		t.Errorf("rescue hook args = %v, want {ws-stuck i-0badf00d provision_timeout_sweep}", got)
+	}
+}
+
+// TestSweep_RescueDoesNotFireOnRace — affected==0 means the row raced to
+// online/restart between SELECT and UPDATE. That is NOT a boot-failure
+// verdict, so the rescue capture must NOT fire (we'd be snapshotting a
+// healthy box that's about to come online).
+func TestSweep_RescueDoesNotFireOnRace(t *testing.T) {
+	mock := setupTestDB(t)
+	rec := withRescueHook(t)
+
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-raced", "codex", "i-raced", 700}))
+	mock.ExpectExec(`UPDATE workspaces`).
+		WithArgs("ws-raced", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
+		WillReturnResult(sqlmock.NewResult(0, 0)) // raced — 0 rows
+
+	sweepStuckProvisioning(context.Background(), &fakeEmitter{}, nil)
+
+	if rec.count() != 0 {
+		t.Errorf("rescue hook must NOT fire on a raced flip (affected==0), got %d calls", rec.count())
+	}
+}
+
+// TestSweep_RescueDoesNotFireOnHealthyRow — a not-yet-overdue row is
+// never flipped, so the rescue capture must not fire. Guards against the
+// hook being attached above the age gate.
+func TestSweep_RescueDoesNotFireOnHealthyRow(t *testing.T) {
+	mock := setupTestDB(t)
+	rec := withRescueHook(t)
+
+	// hermes at 11 min (660s) < 30 min hermes budget → not overdue, no flip.
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-healthy", "hermes", "i-healthy", 660}))
+
+	sweepStuckProvisioning(context.Background(), &fakeEmitter{}, nil)
+
+	if rec.count() != 0 {
+		t.Errorf("rescue hook must NOT fire on a non-overdue (healthy) row, got %d calls", rec.count())
+	}
+}
+
+// TestSweep_RescueNilHookIsSafe — on a deploy where the hook is unwired
+// (self-hosted / no rescue shipping), the sweep must still flip + emit
+// without panicking on the nil hook.
+func TestSweep_RescueNilHookIsSafe(t *testing.T) {
+	mock := setupTestDB(t)
+	prev := BootFailureRescueHook
+	BootFailureRescueHook = nil
+	t.Cleanup(func() { BootFailureRescueHook = prev })
+
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-stuck", "codex", "i-x", 700}))
+	mock.ExpectExec(`UPDATE workspaces`).
+		WithArgs("ws-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	emit := &fakeEmitter{}
+	sweepStuckProvisioning(context.Background(), emit, nil) // must not panic
+
+	if emit.count() != 1 {
+		t.Errorf("flip+emit must still happen with a nil rescue hook, got %d events", emit.count())
+	}
+}
@@ -7,8 +7,8 @@ import (
 	"testing"
 	"time"

-	"github.com/DATA-DOG/go-sqlmock"
 	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/models"
+	"github.com/DATA-DOG/go-sqlmock"
 )

 // fakeEmitter records every RecordAndBroadcast call so tests can assert
@@ -42,12 +42,15 @@ func (f *fakeEmitter) count() int {
 	return len(f.events)
 }

-// candidateRows builds the new-shape query result (id, runtime, age_sec).
-// Use this in every sweep test to match the runtime-aware SELECT.
-func candidateRows(rows ...[3]any) *sqlmock.Rows {
-	r := sqlmock.NewRows([]string{"id", "runtime", "age_sec"})
+// candidateRows builds the query result (id, runtime, instance_id,
+// age_sec). instance_id was added for the RFC internal#742 rescue hook —
+// it rides alongside runtime so the boot-failure capture can reach the
+// still-running box. Tests that don't care about the rescue path pass
+// "" for instance_id. Use this in every sweep test to match the SELECT.
+func candidateRows(rows ...[4]any) *sqlmock.Rows {
+	r := sqlmock.NewRows([]string{"id", "runtime", "instance_id", "age_sec"})
 	for _, row := range rows {
-		r = r.AddRow(row[0], row[1], row[2])
+		r = r.AddRow(row[0], row[1], row[2], row[3])
 	}
 	return r
 }
@@ -58,8 +61,8 @@ func TestSweepStuckProvisioning_FlipsOverdue(t *testing.T) {
 	mock := setupTestDB(t)

 	// claude-code workspace, 700s old > 600s default timeout → flipped.
-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-stuck", "claude-code", 700}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-stuck", "claude-code", "i-stuck", 700}))

 	mock.ExpectExec(`UPDATE workspaces`).
 		WithArgs("ws-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
@@ -92,8 +95,8 @@ func TestSweepStuckProvisioning_HermesGets30MinSlack(t *testing.T) {

 	// 11 min = 660 sec. < HermesProvisioningTimeout (1800s).
 	// No UPDATE should fire — hermes still has time.
-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-hermes-booting", "hermes", 660}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-hermes-booting", "hermes", "i-h1", 660}))

 	emit := &fakeEmitter{}
 	sweepStuckProvisioning(context.Background(), emit, nil)
@@ -114,8 +117,8 @@ func TestSweepStuckProvisioning_HermesPastDeadline(t *testing.T) {
 	mock := setupTestDB(t)

 	// 31 min = 1860 sec > HermesProvisioningTimeout (1800s).
-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-hermes-stuck", "hermes", 1860}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-hermes-stuck", "hermes", "i-h2", 1860}))
 	mock.ExpectExec(`UPDATE workspaces`).
 		WithArgs("ws-hermes-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
 		WillReturnResult(sqlmock.NewResult(0, 1))
@@ -150,8 +153,8 @@ func TestSweepStuckProvisioning_HermesPastDeadline(t *testing.T) {
 func TestSweepStuckProvisioning_ManifestOverrideSparesRow(t *testing.T) {
 	mock := setupTestDB(t)

-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-claude-templated", "claude-code", 660}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-claude-templated", "claude-code", "i-ct", 660}))

 	// No ExpectExec — if the sweeper still flips the row, sqlmock will
 	// fail with an unexpected-query error.
@@ -183,8 +186,8 @@ func TestSweepStuckProvisioning_ManifestOverrideStillFlipsPastDeadline(t *testin
 	mock := setupTestDB(t)

 	// 21 min = 1260s > 1200s manifest override → flipped.
-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-claude-truly-stuck", "claude-code", 1260}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-claude-truly-stuck", "claude-code", "i-cts", 1260}))
 	mock.ExpectExec(`UPDATE workspaces`).
 		WithArgs("ws-claude-truly-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
 		WillReturnResult(sqlmock.NewResult(0, 1))
@@ -221,8 +224,8 @@ func TestSweepStuckProvisioning_ManifestOverrideStillFlipsPastDeadline(t *testin
 func TestSweepStuckProvisioning_RaceSafe(t *testing.T) {
 	mock := setupTestDB(t)

-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-raced", "claude-code", 700}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-raced", "claude-code", "i-raced", 700}))

 	mock.ExpectExec(`UPDATE workspaces`).
 		WithArgs("ws-raced", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
@@ -244,7 +247,7 @@ func TestSweepStuckProvisioning_RaceSafe(t *testing.T) {
 func TestSweepStuckProvisioning_NoStuck(t *testing.T) {
 	mock := setupTestDB(t)

-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
 		WillReturnRows(candidateRows())

 	emit := &fakeEmitter{}
@@ -265,10 +268,10 @@ func TestSweepStuckProvisioning_NoStuck(t *testing.T) {
 func TestSweepStuckProvisioning_MultipleStuck(t *testing.T) {
 	mock := setupTestDB(t)

-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
 		WillReturnRows(candidateRows(
-			[3]any{"ws-claude-code", "claude-code", 700},
-			[3]any{"ws-hermes", "hermes", 1860},
+			[4]any{"ws-claude-code", "claude-code", "i-cc", 700},
+			[4]any{"ws-hermes", "hermes", "i-hh", 1860},
 		))

 	mock.ExpectExec(`UPDATE workspaces`).
@@ -292,8 +295,8 @@ func TestSweepStuckProvisioning_MultipleStuck(t *testing.T) {
 func TestSweepStuckProvisioning_BroadcastFailureDoesNotCrash(t *testing.T) {
 	mock := setupTestDB(t)

-	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), EXTRACT`).
-		WillReturnRows(candidateRows([3]any{"ws-stuck", "claude-code", 700}))
+	mock.ExpectQuery(`SELECT id, COALESCE\(runtime, ''\), COALESCE\(instance_id, ''\), EXTRACT`).
+		WillReturnRows(candidateRows([4]any{"ws-stuck", "claude-code", "i-stuck", 700}))
 	mock.ExpectExec(`UPDATE workspaces`).
 		WithArgs("ws-stuck", sqlmock.AnyArg(), sqlmock.AnyArg(), models.StatusFailed).
 		WillReturnResult(sqlmock.NewResult(0, 1))
@@ -0,0 +1,321 @@
+// Package rescue captures a fixed post-mortem "rescue bundle" off a
+// workspace EC2 whose boot FAILED — before the platform's sweeper /
+// control-plane reaps the instance — and ships it to obs/Loki so a
+// wedged workspace (e.g. the codex provider-derivation failure that
+// motivated RFC internal#742) is inspectable instead of an
+// uninspectable wall.
+//
+// Design constraints (RFC internal#742, Part 2):
+//
+//   - BEST-EFFORT + NON-BLOCKING. Capture MUST NOT change boot-failure
+//     semantics or add latency to the failure path. Callers fire
+//     Capture in its own goroutine; Capture additionally bounds itself
+//     with CaptureTimeout so a hung EIC tunnel can't wedge the
+//     goroutine forever.
+//   - FIRES ON THE BOOT-FAILURE VERDICT ONLY. The two hook points are
+//     the provision-timeout sweep (registry.sweepStuckProvisioning) and
+//     the out-of-band bootstrap-watcher signal
+//     (handlers.WorkspaceHandler.BootstrapFailed). Normal teardown /
+//     deprovision / recreate / billing-suspend / hibernate paths do NOT
+//     call Capture — see the RFC's path enumeration.
+//   - REDACT BEFORE ANYTHING LEAVES THE BOX. Every collected section is
+//     run through the injected Redact func (wired to the existing
+//     handlers.redactSecrets secret-scan) before it is shipped. Raw
+//     tokens/keys never reach Loki.
+//
+// The package is a LEAF: it imports only internal/audit (the obs
+// shipper) so it can be called from both handlers and registry without
+// an import cycle (registry must not import handlers). The two heavy
+// dependencies — the EIC/SSH remote-command runner and the redactor —
+// are injected as package-level func vars, wired once at boot from the
+// handlers package (which owns withEICTunnel + redactSecrets). Tests
+// swap them for fakes.
+package rescue
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/audit"
+)
+
+// CaptureTimeout bounds the whole bundle collection. The sweeper runs
+// every 30s and the CP reap follows the failure verdict; 45s gives the
+// EIC dance (~3-5s) plus six short remote commands (<2s each) generous
+// headroom while still finishing well before the instance is torn down.
+// Distinct from the per-op eicFileOpTimeout so a slow box that already
+// failed to boot can't hang the capture goroutine indefinitely.
+const CaptureTimeout = 45 * time.Second
+
+// LokiKind is the Loki stream label value that tags every rescue
+// record. Queryable as `kind="rescue"` (RFC internal#742 §Loki labels).
+const LokiKind = "rescue"
+
+// RescueVolumeGrace is how long a boot-failed workspace's /configs data
+// volume (and its still-running instance) must be RETAINED past the
+// boot-failure verdict so a live rescue read is possible — distinct from
+// the user-requested prune path (cp#415), which is an explicit erase.
+//
+// In molecule-core (the tenant platform) the boot-failure verdict only
+// flips workspaces.status to `failed`; it never issues a terminate. The
+// platform's two reapers (registry.StartCPOrphanSweeper +
+// handlers deprovision) act ONLY on status='removed', so a `failed`
+// workspace's instance + /configs volume are retained here by
+// construction — see TestCPSweepOnce_DoesNotReapFailedWorkspace. The
+// time-bounded reap of the failed instance is the control plane's
+// bootstrap-watcher concern; this constant is the SSOT for the grace
+// the CP must honour (24h covers an operator's next-business-day
+// post-mortem without leaking the volume indefinitely).
+const RescueVolumeGrace = 24 * time.Hour
+
+// rescueEventType is the audit event_type carried in the shipped
+// record. The obs shipper (internal/audit) already maps event_type to a
+// low-cardinality Loki label; "rescue.bundle" keeps the rescue stream
+// trivially filterable alongside the existing audit taxonomy.
+const rescueEventType = "rescue.bundle"
+
+// RunRemote runs a single shell command on the still-running (but
+// unconfigured) workspace EC2 over EIC/SSH and returns its combined
+// output. Wired at boot to the handlers EIC runner
+// (rescueRunRemoteViaEIC). nil until wired — Capture degrades to a
+// logged no-op rather than panicking, so an operator who hasn't wired
+// the hook still gets a clear signal instead of a crash on the failure
+// path.
+var RunRemote func(ctx context.Context, instanceID, command string) (string, error)
+
+// Redact scrubs secret-shaped substrings from a collected section
+// before it leaves the box. Wired at boot to handlers.redactSecrets.
+// nil until wired — Capture refuses to ship un-redacted content if the
+// redactor is missing (fails closed: logs + aborts rather than leaking
+// raw config).
+var Redact func(workspaceID, content string) string
+
+// section is one labelled chunk of the rescue bundle: a human-readable
+// name + the remote command that produces it.
+type section struct {
+	name    string
+	command string
+}
+
+// bundleSections is the FIXED set collected on every boot-failure
+// rescue (RFC internal#742 §Build.1). Order is the post-mortem reading
+// order: config first, then boot logs, then container state, then the
+// resolved model/provider env that drove the codex derivation failure.
+//
+//   - /configs/config.yaml + system-prompt.md: the managed config the
+//     runtime booted against (redacted; system-prompt can embed keys).
+//   - cloud-init-output.log tail: the user-data execution trace — where
+//     a wedged boot actually died.
+//   - docker ps -a: container state (did the agent container even
+//     start, exit-code, restart loop).
+//   - agent container logs: the runtime's own stderr (the codex
+//     provider-derivation panic lives here).
+//   - MODEL|PROVIDER|RUNTIME env: the resolved routing that motivated
+//     the RFC. `sudo cat` of the container env via docker inspect-style
+//     grep — see the command.
+//
+// All commands use `sudo -n` (the box's /configs is root-owned; ubuntu
+// has passwordless sudo) and swallow missing-target stderr so a section
+// that can't be produced ships as a short marker instead of failing the
+// whole bundle. Kept as data (not inlined) so the redaction + ship loop
+// is uniform and the set is reviewable in one place.
+var bundleSections = []section{
+	{
+		name:    "config.yaml",
+		command: "sudo -n cat /configs/config.yaml 2>/dev/null || echo '(/configs/config.yaml absent)'",
+	},
+	{
+		name:    "system-prompt.md",
+		command: "sudo -n cat /configs/system-prompt.md 2>/dev/null || echo '(/configs/system-prompt.md absent)'",
+	},
+	{
+		name:    "cloud-init-output.log.tail",
+		command: "sudo -n tail -200 /var/log/cloud-init-output.log 2>/dev/null || echo '(cloud-init-output.log absent)'",
+	},
+	{
+		name:    "docker-ps",
+		command: "sudo -n docker ps -a 2>/dev/null || echo '(docker unavailable)'",
+	},
+	{
+		// The agent container is the first non-infra container; grab the
+		// most recently created one and tail its logs. `head -1` of
+		// `docker ps -a -q` is creation-ordered newest-first, which is
+		// the agent runtime on a workspace box.
+		name:    "agent-container.logs.tail",
+		command: "cid=$(sudo -n docker ps -a -q 2>/dev/null | head -1); [ -n \"$cid\" ] && sudo -n docker logs --tail 200 \"$cid\" 2>&1 || echo '(no agent container)'",
+	},
+	{
+		// Resolved model/provider/runtime env from the agent container.
+		// `docker inspect` the env array and grep the routing keys. This
+		// is the field that pinpoints a provider-derivation failure.
+		name:    "model-provider-runtime.env",
+		command: "cid=$(sudo -n docker ps -a -q 2>/dev/null | head -1); [ -n \"$cid\" ] && sudo -n docker inspect --format '{{range .Config.Env}}{{println .}}{{end}}' \"$cid\" 2>/dev/null | grep -E 'MODEL|PROVIDER|RUNTIME' || echo '(no env)'",
+	},
+}
+
+// Input is the identity of the failed workspace being rescued.
+type Input struct {
+	InstanceID  string // EC2 instance id of the still-running failed box
+	WorkspaceID string
+	OrgID       string
+	// Reason is a short tag for WHY the rescue fired (e.g.
+	// "provision_timeout_sweep" or "bootstrap_watcher") — carried into
+	// the Loki record so an operator can correlate the bundle with the
+	// failure verdict that triggered it.
+	Reason string
+}
+
+// Section is one labelled, already-redacted chunk of the persisted
+// rescue bundle. It mirrors what ship() emits to Loki per-section, but
+// is the unit the queryable store (and the read endpoint) returns.
+// `Redacted` is false only for collection-failure markers (the section
+// command couldn't run); true sections passed through the secret-scan.
+type Section struct {
+	Name     string `json:"name"`
+	Content  string `json:"content"`
+	Redacted bool   `json:"redacted"`
+}
+
+// Bundle is the full, already-redacted post-mortem bundle for ONE
+// boot-failure capture — the unit persisted to the queryable store on
+// capture (RFC internal#742 Part 3) and served by
+// GET /workspaces/:id/rescue. Sections are in fixed reading order
+// (config → boot logs → container state → resolved routing env).
+type Bundle struct {
+	WorkspaceID string    `json:"workspace_id"`
+	OrgID       string    `json:"org_id"`
+	InstanceID  string    `json:"instance_id"`
+	Reason      string    `json:"reason"`
+	Sections    []Section `json:"sections"`
+}
+
+// PersistBundle writes the fully-collected, already-redacted bundle to
+// the queryable per-tenant store (rescue_bundles table) so the rescue
+// READ endpoint can serve it without obs/Loki read creds (RFC
+// internal#742 Part 3 read-path decision — see the migration header).
+//
+// Wired at boot from the handlers package (which owns db.DB) to keep
+// internal/rescue a leaf: it must NOT import internal/db, or registry —
+// which imports rescue — would inherit a db dependency it can call
+// without a cycle, but more importantly the leaf stays trivially
+// testable with a fake. nil until wired: a capture with no store wired
+// still ships to Loki (Part 2 behavior preserved) and logs that it
+// skipped the DB persist, rather than failing the capture.
+var PersistBundle func(ctx context.Context, b Bundle) error
+
+// Capture collects the fixed rescue bundle off the failed instance,
+// redacts each section, and ships it to Loki under
+// {kind="rescue", org=<OrgID>, workspace_id=<WorkspaceID>}.
+//
+// BEST-EFFORT: every failure mode (missing wiring, EIC error, a single
+// section that won't collect) is logged and does NOT propagate — Capture
+// never returns an error and never panics, so the boot-failure handling
+// at the call site is unaffected. The caller is expected to invoke this
+// in its own goroutine; Capture additionally self-bounds with
+// CaptureTimeout.
+func Capture(ctx context.Context, in Input) {
+	defer func() {
+		// A logging helper on the failure path must never take the
+		// process down. Recover defensively — the redactor / shipper are
+		// injected and a future mis-wire shouldn't crash the sweeper.
+		if r := recover(); r != nil {
+			log.Printf("rescue: capture panicked for ws=%s instance=%s: %v", in.WorkspaceID, in.InstanceID, r)
+		}
+	}()
+
+	if in.InstanceID == "" {
+		// No live box to read — nothing to rescue (e.g. failure before
+		// any EC2 was launched). Not an error; just skip.
+		log.Printf("rescue: skip ws=%s — no instance_id (nothing to capture)", in.WorkspaceID)
+		return
+	}
+	if RunRemote == nil {
+		log.Printf("rescue: skip ws=%s instance=%s — RunRemote not wired (best-effort no-op)", in.WorkspaceID, in.InstanceID)
+		return
+	}
+	if Redact == nil {
+		// Fail CLOSED: without a redactor we could leak raw tokens to
+		// Loki. Abort rather than ship unredacted.
+		log.Printf("rescue: ABORT ws=%s instance=%s — Redact not wired; refusing to ship un-redacted bundle", in.WorkspaceID, in.InstanceID)
+		return
+	}
+
+	ctx, cancel := context.WithTimeout(ctx, CaptureTimeout)
+	defer cancel()
+
+	log.Printf("rescue: capturing bundle ws=%s instance=%s reason=%s", in.WorkspaceID, in.InstanceID, in.Reason)
+
+	collected := 0
+	// Accumulate the per-section result alongside shipping each to Loki,
+	// so the same already-redacted content is persisted to the queryable
+	// store as one bundle row after the loop. Shipping stays per-section
+	// (Part 2 Loki behavior unchanged); persistence is the single
+	// bundle the read endpoint serves.
+	bundle := Bundle{
+		WorkspaceID: in.WorkspaceID,
+		OrgID:       in.OrgID,
+		InstanceID:  in.InstanceID,
+		Reason:      in.Reason,
+		Sections:    make([]Section, 0, len(bundleSections)),
+	}
+	for _, sec := range bundleSections {
+		raw, err := RunRemote(ctx, in.InstanceID, sec.command)
+		if err != nil {
+			// One section failing (e.g. ssh blip mid-collection) must not
+			// abort the rest — ship a marker for it and continue.
+			log.Printf("rescue: section %q failed for ws=%s: %v", sec.name, in.WorkspaceID, err)
+			marker := fmt.Sprintf("(rescue: section collection failed: %v)", err)
+			ship(ctx, in, sec.name, marker, false)
+			bundle.Sections = append(bundle.Sections, Section{Name: sec.name, Content: marker, Redacted: false})
+			continue
+		}
+		redacted := Redact(in.WorkspaceID, raw)
+		ship(ctx, in, sec.name, redacted, true)
+		bundle.Sections = append(bundle.Sections, Section{Name: sec.name, Content: redacted, Redacted: true})
+		collected++
+	}
+
+	log.Printf("rescue: shipped %d/%d sections ws=%s instance=%s kind=%s", collected, len(bundleSections), in.WorkspaceID, in.InstanceID, LokiKind)
+
+	// Persist the redacted bundle to the queryable store so the rescue
+	// READ endpoint can serve it without obs/Loki read creds. Best-effort
+	// and last: a persist failure (or no store wired) must NOT undo the
+	// Loki ship that already succeeded, and never panics the failure path.
+	persistBundle(ctx, bundle)
+}
+
+// persistBundle writes the collected bundle to the queryable store if a
+// store is wired. Best-effort: a nil store (operator hasn't wired the
+// READ path) or a DB error is logged and swallowed — the Loki ship is
+// the durable cross-tenant copy, and the failure path must never be
+// disturbed by the post-mortem read store.
+func persistBundle(ctx context.Context, b Bundle) {
+	if PersistBundle == nil {
+		log.Printf("rescue: store not wired — bundle for ws=%s shipped to Loki only (no queryable copy)", b.WorkspaceID)
+		return
+	}
+	if err := PersistBundle(ctx, b); err != nil {
+		log.Printf("rescue: persist bundle for ws=%s failed (shipped to Loki regardless): %v", b.WorkspaceID, err)
+	}
+}
+
+// ship emits one rescue section to Loki via the audit shipper. The
+// org / workspace_id / kind ride in the record body (queryable via
+// LogQL `| json`); event_type ("rescue.bundle") is the low-cardinality
+// Loki label the shipper already promotes. `redacted` records whether
+// the content passed through the secret-scan, so an operator can tell a
+// shipped-but-redacted section from a collection-failure marker.
+func ship(ctx context.Context, in Input, name, content string, redacted bool) {
+	audit.Emit(ctx, rescueEventType, map[string]any{
+		"kind":         LokiKind,
+		"org":          in.OrgID,
+		"workspace_id": in.WorkspaceID,
+		"instance_id":  in.InstanceID,
+		"reason":       in.Reason,
+		"section":      name,
+		"redacted":     redacted,
+		"content":      content,
+	})
+}
@@ -0,0 +1,136 @@
+package rescue
+
+// Part 3 coverage: Capture, after collecting + redacting every section,
+// persists the bundle exactly once to the queryable store (in addition
+// to the per-section Loki ship verified in rescue_test.go).
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+)
+
+// withPersist swaps the injected PersistBundle for the test and restores
+// it after.
+func withPersist(t *testing.T, fn func(ctx context.Context, b Bundle) error) {
+	t.Helper()
+	prev := PersistBundle
+	PersistBundle = fn
+	t.Cleanup(func() { PersistBundle = prev })
+}
+
+// TestCapture_PersistsBundleOnce: the happy path persists one bundle
+// carrying every section, with identity + redacted content matching what
+// was shipped.
+func TestCapture_PersistsBundleOnce(t *testing.T) {
+	_ = captureLoki(t) // keep Loki transport pointed at a temp file
+	withFakes(t,
+		func(_ context.Context, instanceID, cmd string) (string, error) {
+			return "OUT:" + instanceID, nil
+		},
+		func(_ws, c string) string { return "RED:" + c },
+	)
+
+	var persisted []Bundle
+	withPersist(t, func(_ context.Context, b Bundle) error {
+		persisted = append(persisted, b)
+		return nil
+	})
+
+	Capture(context.Background(), Input{
+		InstanceID:  "i-abc",
+		WorkspaceID: "ws-1",
+		OrgID:       "org-9",
+		Reason:      "provision_timeout_sweep",
+	})
+
+	if len(persisted) != 1 {
+		t.Fatalf("PersistBundle called %d times, want exactly 1", len(persisted))
+	}
+	b := persisted[0]
+	if b.WorkspaceID != "ws-1" || b.OrgID != "org-9" || b.InstanceID != "i-abc" || b.Reason != "provision_timeout_sweep" {
+		t.Errorf("bundle identity wrong: %+v", b)
+	}
+	if len(b.Sections) != len(bundleSections) {
+		t.Fatalf("persisted %d sections, want %d", len(b.Sections), len(bundleSections))
+	}
+	for _, s := range b.Sections {
+		if !s.Redacted {
+			t.Errorf("section %q persisted with redacted=false on the happy path", s.Name)
+		}
+		// Redactor ("RED:" prefix) must have run on persisted content.
+		if !strings.HasPrefix(s.Content, "RED:") {
+			t.Errorf("section %q persisted un-redacted content: %q", s.Name, s.Content)
+		}
+	}
+}
+
+// TestCapture_PersistFailureDoesNotPanic: a store error is swallowed —
+// Capture still completes (the Loki ship already succeeded).
+func TestCapture_PersistFailureDoesNotPanic(t *testing.T) {
+	_ = captureLoki(t)
+	withFakes(t,
+		func(_ context.Context, _ string, _ string) (string, error) { return "ok", nil },
+		func(_ws, c string) string { return c },
+	)
+	withPersist(t, func(_ context.Context, _ Bundle) error {
+		return errors.New("db down")
+	})
+	// Must not panic / must return normally.
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-2", OrgID: "o"})
+}
+
+// TestCapture_NoPersistWiredIsSafe: with PersistBundle unwired (operator
+// hasn't wired the read path), Capture still ships to Loki and does not
+// panic.
+func TestCapture_NoPersistWiredIsSafe(t *testing.T) {
+	readLoki := captureLoki(t)
+	withFakes(t,
+		func(_ context.Context, _ string, _ string) (string, error) { return "ok", nil },
+		func(_ws, c string) string { return c },
+	)
+	prev := PersistBundle
+	PersistBundle = nil
+	t.Cleanup(func() { PersistBundle = prev })
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-3", OrgID: "o"})
+
+	// Loki ship still happened for every section.
+	if recs := readLoki(); len(recs) != len(bundleSections) {
+		t.Errorf("shipped %d records, want %d (Loki unaffected by missing store)", len(recs), len(bundleSections))
+	}
+}
+
+// TestCapture_FailureMarkerPersistedAsNonRedacted: a section whose
+// collection fails is persisted with redacted=false + a marker, matching
+// the Loki record.
+func TestCapture_FailureMarkerPersistedAsNonRedacted(t *testing.T) {
+	_ = captureLoki(t)
+	withFakes(t,
+		func(_ context.Context, _ string, cmd string) (string, error) {
+			if strings.Contains(cmd, "config.yaml") {
+				return "", errors.New("ssh blip")
+			}
+			return "ok", nil
+		},
+		func(_ws, c string) string { return c },
+	)
+	var got Bundle
+	withPersist(t, func(_ context.Context, b Bundle) error { got = b; return nil })
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-4", OrgID: "o"})
+
+	var markers int
+	for _, s := range got.Sections {
+		if !s.Redacted {
+			markers++
+			if !strings.Contains(s.Content, "section collection failed") {
+				t.Errorf("non-redacted section %q content = %q, want a failure marker", s.Name, s.Content)
+			}
+		}
+	}
+	if markers != 1 {
+		t.Errorf("want exactly 1 failure marker persisted, got %d", markers)
+	}
+}
@@ -0,0 +1,226 @@
+package rescue
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// withFakes swaps the injected RunRemote + Redact for the duration of a
+// test and restores them after. Mirrors the provisioner test-fake
+// pattern (package-var swap + t.Cleanup).
+func withFakes(t *testing.T, run func(ctx context.Context, instanceID, cmd string) (string, error), redact func(ws, c string) string) {
+	t.Helper()
+	prevRun, prevRedact := RunRemote, Redact
+	RunRemote = run
+	Redact = redact
+	t.Cleanup(func() { RunRemote = prevRun; Redact = prevRedact })
+}
+
+// captureLoki points the audit shipper at a temp JSONL file and returns
+// a reader that decodes the records the rescue ship() loop wrote. This
+// is the same transport the production rescue stream uses (audit.Emit →
+// Loki via the tenant Vector source), so asserting on it proves the
+// shipper-reuse + labels end to end.
+func captureLoki(t *testing.T) func() []map[string]any {
+	t.Helper()
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.jsonl")
+	t.Setenv("MOLECULE_AUDIT_LOG_PATH", path)
+	return func() []map[string]any {
+		b, err := os.ReadFile(path)
+		if err != nil {
+			return nil
+		}
+		var out []map[string]any
+		for _, line := range strings.Split(strings.TrimSpace(string(b)), "\n") {
+			if line == "" {
+				continue
+			}
+			var rec map[string]any
+			if err := json.Unmarshal([]byte(line), &rec); err != nil {
+				t.Fatalf("bad audit jsonl line %q: %v", line, err)
+			}
+			out = append(out, rec)
+		}
+		return out
+	}
+}
+
+func fields(rec map[string]any) map[string]any {
+	f, _ := rec["fields"].(map[string]any)
+	return f
+}
+
+// TestCapture_ShipsAllSectionsWithRescueLabels is the happy path: a
+// boot-failure capture collects every fixed section, runs each through
+// the redactor, and ships it to Loki under {kind="rescue", org, ws}.
+func TestCapture_ShipsAllSectionsWithRescueLabels(t *testing.T) {
+	readLoki := captureLoki(t)
+	var seenCmds []string
+	withFakes(t,
+		func(_ context.Context, instanceID, cmd string) (string, error) {
+			seenCmds = append(seenCmds, cmd)
+			return "OUTPUT for " + instanceID, nil
+		},
+		func(_ws, c string) string { return c }, // identity redactor
+	)
+
+	Capture(context.Background(), Input{
+		InstanceID:  "i-abc123",
+		WorkspaceID: "ws-1",
+		OrgID:       "org-9",
+		Reason:      "provision_timeout_sweep",
+	})
+
+	recs := readLoki()
+	if len(recs) != len(bundleSections) {
+		t.Fatalf("want %d shipped sections, got %d", len(bundleSections), len(recs))
+	}
+	if len(seenCmds) != len(bundleSections) {
+		t.Fatalf("want %d remote commands run, got %d", len(bundleSections), len(seenCmds))
+	}
+	for _, rec := range recs {
+		if rec["event_type"] != rescueEventType {
+			t.Errorf("event_type = %v, want %q", rec["event_type"], rescueEventType)
+		}
+		// workspace_id is promoted to the top-level record position by
+		// the audit shipper.
+		if rec["workspace_id"] != "ws-1" {
+			t.Errorf("top-level workspace_id = %v, want ws-1", rec["workspace_id"])
+		}
+		f := fields(rec)
+		if f["kind"] != LokiKind {
+			t.Errorf("kind = %v, want %q", f["kind"], LokiKind)
+		}
+		if f["org"] != "org-9" {
+			t.Errorf("org = %v, want org-9", f["org"])
+		}
+		if f["instance_id"] != "i-abc123" {
+			t.Errorf("instance_id = %v, want i-abc123", f["instance_id"])
+		}
+		if f["redacted"] != true {
+			t.Errorf("redacted = %v, want true for a collected section", f["redacted"])
+		}
+	}
+}
+
+// TestCapture_Redacts proves the bundle is scrubbed before it leaves the
+// box: a remote section that contains a secret-shaped token must ship
+// with the token replaced, never raw.
+func TestCapture_Redacts(t *testing.T) {
+	readLoki := captureLoki(t)
+	const secret = "sk-ant-SUPERSECRETTOKENVALUE0001"
+	withFakes(t,
+		func(_ context.Context, _ string, _ string) (string, error) {
+			return "ANTHROPIC_API_KEY=" + secret, nil
+		},
+		// redactor that mangles anything containing the secret shape
+		func(_ws, c string) string {
+			if strings.Contains(c, secret) {
+				return strings.ReplaceAll(c, secret, "[REDACTED]")
+			}
+			return c
+		},
+	)
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-2", OrgID: "o"})
+
+	for _, rec := range readLoki() {
+		content, _ := fields(rec)["content"].(string)
+		if strings.Contains(content, secret) {
+			t.Fatalf("raw secret leaked to Loki in section %v: %q", fields(rec)["section"], content)
+		}
+	}
+}
+
+// TestCapture_SkipsWhenNoInstance: a failure with no provisioned EC2 has
+// nothing to read — Capture must no-op (ship nothing) rather than dial a
+// blank instance id.
+func TestCapture_SkipsWhenNoInstance(t *testing.T) {
+	readLoki := captureLoki(t)
+	called := false
+	withFakes(t,
+		func(_ context.Context, _ string, _ string) (string, error) { called = true; return "", nil },
+		func(_ws, c string) string { return c },
+	)
+	Capture(context.Background(), Input{InstanceID: "", WorkspaceID: "ws-3", OrgID: "o"})
+	if called {
+		t.Error("RunRemote called for an empty instance id")
+	}
+	if recs := readLoki(); len(recs) != 0 {
+		t.Errorf("shipped %d records for an empty instance id, want 0", len(recs))
+	}
+}
+
+// TestCapture_FailsClosedWithoutRedactor: if the redactor is not wired,
+// Capture must NOT ship anything (would leak raw config). Fail closed.
+func TestCapture_FailsClosedWithoutRedactor(t *testing.T) {
+	readLoki := captureLoki(t)
+	prevRun, prevRedact := RunRemote, Redact
+	RunRemote = func(_ context.Context, _ string, _ string) (string, error) { return "raw config", nil }
+	Redact = nil
+	t.Cleanup(func() { RunRemote = prevRun; Redact = prevRedact })
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-4", OrgID: "o"})
+
+	if recs := readLoki(); len(recs) != 0 {
+		t.Errorf("shipped %d records without a redactor wired, want 0 (fail closed)", len(recs))
+	}
+}
+
+// TestCapture_SectionFailureIsIsolated: one section's RunRemote error
+// must not abort the rest — the failing section ships a marker and the
+// others still ship.
+func TestCapture_SectionFailureIsIsolated(t *testing.T) {
+	readLoki := captureLoki(t)
+	withFakes(t,
+		func(_ context.Context, _ string, cmd string) (string, error) {
+			if strings.Contains(cmd, "config.yaml") {
+				return "", errors.New("ssh blip")
+			}
+			return "ok", nil
+		},
+		func(_ws, c string) string { return c },
+	)
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-5", OrgID: "o"})
+
+	recs := readLoki()
+	if len(recs) != len(bundleSections) {
+		t.Fatalf("want all %d sections shipped (incl. failure marker), got %d", len(bundleSections), len(recs))
+	}
+	var failureMarkers int
+	for _, rec := range recs {
+		if fields(rec)["redacted"] == false {
+			failureMarkers++
+			content, _ := fields(rec)["content"].(string)
+			if !strings.Contains(content, "section collection failed") {
+				t.Errorf("failure marker content = %q, want a collection-failed marker", content)
+			}
+		}
+	}
+	if failureMarkers != 1 {
+		t.Errorf("want exactly 1 failure marker, got %d", failureMarkers)
+	}
+}
+
+// TestCapture_NoWiringIsSafeNoOp: with RunRemote unwired (operator hasn't
+// called the boot wiring), Capture must be a logged no-op, never a panic.
+func TestCapture_NoWiringIsSafeNoOp(t *testing.T) {
+	readLoki := captureLoki(t)
+	prevRun, prevRedact := RunRemote, Redact
+	RunRemote = nil
+	Redact = func(_ws, c string) string { return c }
+	t.Cleanup(func() { RunRemote = prevRun; Redact = prevRedact })
+
+	Capture(context.Background(), Input{InstanceID: "i-x", WorkspaceID: "ws-6", OrgID: "o"})
+
+	if recs := readLoki(); len(recs) != 0 {
+		t.Errorf("shipped %d records with RunRemote unwired, want 0", len(recs))
+	}
+}
@@ -0,0 +1,153 @@
+// Package rescuestore is the queryable persistence layer for rescue
+// bundles (RFC internal#742 Part 3). It is the DB side of the read-path
+// decision: because internal/audit (Part 2's ship transport) is
+// Loki-only and tenants hold no obs read creds, the redacted bundle is
+// ALSO written here on capture so GET /workspaces/:id/rescue can serve
+// the latest one with a plain Postgres read.
+//
+// The package owns both the write (Persist, wired into
+// rescue.PersistBundle at boot) and the read (GetLatest, used by the
+// handler). It depends on internal/db and internal/rescue (for the
+// Bundle/Section types); it is imported by handlers, never by the leaf
+// internal/rescue or by registry — so no import cycle.
+package rescuestore
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
+)
+
+// maxSectionBytes bounds a single persisted section's content so a
+// pathological capture (e.g. a multi-megabyte container log) can't bloat
+// the row or the read response. Capture already tails to ~200 lines per
+// section, so this is a backstop, not the primary limit. Truncated
+// content is suffixed with a marker so a reader knows it was clipped.
+const maxSectionBytes = 64 * 1024 // 64 KiB per section
+
+// truncationMarker is appended to any section clipped at maxSectionBytes.
+const truncationMarker = "\n…(rescue: section truncated at 64KiB)"
+
+// StoredBundle is a persisted bundle plus its capture timestamp (the DB
+// assigns captured_at on write). The handler maps this to the read
+// response shape.
+type StoredBundle struct {
+	Bundle     rescue.Bundle
+	CapturedAt time.Time
+}
+
+// Store is the read/write surface the handler and the capture wiring
+// depend on. An interface so the handler test can fake it without a
+// sqlmock; the production implementation is Postgres.
+type Store interface {
+	// Persist writes one bundle row (captured_at = now()).
+	Persist(ctx context.Context, b rescue.Bundle) error
+	// GetLatest returns the most recent bundle for workspaceID. When
+	// orgID is non-empty the row must also match org_id (cross-org
+	// defense-in-depth behind TenantGuard). Returns (nil, nil) — NOT an
+	// error — when no bundle exists, so the handler can 404 cleanly.
+	GetLatest(ctx context.Context, workspaceID, orgID string) (*StoredBundle, error)
+}
+
+// Postgres is the production Store backed by the rescue_bundles table.
+type Postgres struct{ db *sql.DB }
+
+// NewPostgres builds a Postgres-backed store over the given handle.
+func NewPostgres(db *sql.DB) *Postgres { return &Postgres{db: db} }
+
+// Persist writes the bundle as one row. Sections are stored as JSONB.
+// Each section's content is clamped to maxSectionBytes before write.
+func (p *Postgres) Persist(ctx context.Context, b rescue.Bundle) error {
+	if p.db == nil {
+		return fmt.Errorf("rescuestore: nil db")
+	}
+	clamped := clampSections(b.Sections)
+	payload, err := json.Marshal(clamped)
+	if err != nil {
+		return fmt.Errorf("rescuestore: marshal sections: %w", err)
+	}
+	_, err = p.db.ExecContext(ctx,
+		`INSERT INTO rescue_bundles (workspace_id, org_id, instance_id, reason, sections)
+		 VALUES ($1, $2, $3, $4, $5::jsonb)`,
+		b.WorkspaceID, b.OrgID, b.InstanceID, b.Reason, string(payload),
+	)
+	if err != nil {
+		return fmt.Errorf("rescuestore: insert: %w", err)
+	}
+	return nil
+}
+
+// GetLatest returns the newest bundle for workspaceID, org-scoped. The
+// (workspace_id, captured_at DESC, id DESC) index serves this directly.
+// sql.ErrNoRows maps to (nil, nil) so the handler 404s.
+func (p *Postgres) GetLatest(ctx context.Context, workspaceID, orgID string) (*StoredBundle, error) {
+	if p.db == nil {
+		return nil, fmt.Errorf("rescuestore: nil db")
+	}
+	if orgID == "" {
+		return nil, fmt.Errorf("rescuestore: org_id required")
+	}
+
+	var (
+		instanceID  string
+		reason      string
+		capturedAt  time.Time
+		sectionsRaw []byte
+	)
+	err := p.db.QueryRowContext(ctx,
+		`SELECT instance_id, reason, captured_at, sections
+		   FROM rescue_bundles
+		  WHERE workspace_id = $1
+		    AND org_id = $2
+		  ORDER BY captured_at DESC, id DESC
+		  LIMIT 1`,
+		workspaceID, orgID,
+	).Scan(&instanceID, &reason, &capturedAt, &sectionsRaw)
+	if err == sql.ErrNoRows {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, fmt.Errorf("rescuestore: query latest: %w", err)
+	}
+
+	var sections []rescue.Section
+	if len(sectionsRaw) > 0 {
+		if err := json.Unmarshal(sectionsRaw, &sections); err != nil {
+			return nil, fmt.Errorf("rescuestore: unmarshal sections: %w", err)
+		}
+	}
+
+	return &StoredBundle{
+		Bundle: rescue.Bundle{
+			WorkspaceID: workspaceID,
+			OrgID:       orgID,
+			InstanceID:  instanceID,
+			Reason:      reason,
+			Sections:    sections,
+		},
+		CapturedAt: capturedAt,
+	}, nil
+}
+
+// clampSections returns a copy with each section's content clamped to
+// maxSectionBytes. Clamps on a rune boundary so the marker doesn't split
+// a multibyte sequence — the content is a forensic blob, never parsed.
+func clampSections(in []rescue.Section) []rescue.Section {
+	out := make([]rescue.Section, len(in))
+	for i, s := range in {
+		if len(s.Content) > maxSectionBytes {
+			b := []byte(s.Content[:maxSectionBytes])
+			// Back off to a valid utf-8 boundary (at most 3 bytes).
+			for len(b) > 0 && b[len(b)-1]&0xC0 == 0x80 {
+				b = b[:len(b)-1]
+			}
+			s.Content = string(b) + truncationMarker
+		}
+		out[i] = s
+	}
+	return out
+}
@@ -0,0 +1,211 @@
+package rescuestore
+
+// Sqlmock-backed coverage for the rescue_bundles store (RFC internal#742
+// Part 3). Exercises Persist (incl. section clamp) + GetLatest (happy
+// path, no-rows→nil, org-scoping, query error) without a real DB.
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"errors"
+	"regexp"
+	"strings"
+	"testing"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/rescue"
+	"github.com/DATA-DOG/go-sqlmock"
+)
+
+func newMock(t *testing.T) (*sql.DB, sqlmock.Sqlmock) {
+	t.Helper()
+	dbh, mock, err := sqlmock.New()
+	if err != nil {
+		t.Fatalf("sqlmock: %v", err)
+	}
+	t.Cleanup(func() { _ = dbh.Close() })
+	return dbh, mock
+}
+
+func sampleBundle() rescue.Bundle {
+	return rescue.Bundle{
+		WorkspaceID: "ws-1",
+		OrgID:       "org-9",
+		InstanceID:  "i-abc",
+		Reason:      "bootstrap_watcher",
+		Sections: []rescue.Section{
+			{Name: "config.yaml", Content: "model: gpt-4", Redacted: true},
+			{Name: "docker-ps", Content: "(no agent container)", Redacted: false},
+		},
+	}
+}
+
+// TestPersist_InsertsRow asserts Persist issues one INSERT with the
+// bundle fields and a JSON sections payload.
+func TestPersist_InsertsRow(t *testing.T) {
+	dbh, mock := newMock(t)
+	b := sampleBundle()
+
+	mock.ExpectExec(regexp.QuoteMeta(`INSERT INTO rescue_bundles`)).
+		WithArgs("ws-1", "org-9", "i-abc", "bootstrap_watcher", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+
+	if err := NewPostgres(dbh).Persist(context.Background(), b); err != nil {
+		t.Fatalf("Persist: %v", err)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestClampSections: a section over maxSectionBytes is truncated +
+// marker-suffixed; a small section is untouched.
+func TestClampSections(t *testing.T) {
+	huge := strings.Repeat("x", maxSectionBytes+5000)
+	in := []rescue.Section{
+		{Name: "container.logs", Content: huge, Redacted: true},
+		{Name: "small", Content: "ok", Redacted: true},
+	}
+	out := clampSections(in)
+
+	if len(out[0].Content) > maxSectionBytes+len(truncationMarker) {
+		t.Errorf("clamped content len = %d, want <= %d", len(out[0].Content), maxSectionBytes+len(truncationMarker))
+	}
+	if !strings.HasSuffix(out[0].Content, truncationMarker) {
+		t.Error("clamped section missing truncation marker suffix")
+	}
+	if out[1].Content != "ok" {
+		t.Errorf("small section was modified: %q", out[1].Content)
+	}
+}
+
+// TestPersist_WritesClampedPayload: Persist marshals the clamped
+// sections into the JSONB arg (the INSERT carries the truncation marker).
+func TestPersist_WritesClampedPayload(t *testing.T) {
+	dbh, mock := newMock(t)
+	huge := strings.Repeat("x", maxSectionBytes+5000)
+	b := rescue.Bundle{
+		WorkspaceID: "ws-1",
+		Sections:    []rescue.Section{{Name: "container.logs", Content: huge, Redacted: true}},
+	}
+	want, _ := json.Marshal(clampSections(b.Sections))
+
+	mock.ExpectExec(regexp.QuoteMeta(`INSERT INTO rescue_bundles`)).
+		WithArgs("ws-1", "", "", "", string(want)).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+
+	if err := NewPostgres(dbh).Persist(context.Background(), b); err != nil {
+		t.Fatalf("Persist: %v", err)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet: %v", err)
+	}
+}
+
+// TestGetLatest_ReturnsBundle: a found row decodes back into the bundle.
+func TestGetLatest_ReturnsBundle(t *testing.T) {
+	dbh, mock := newMock(t)
+	ts := time.Date(2026, 5, 31, 12, 0, 0, 0, time.UTC)
+	secs, _ := json.Marshal([]rescue.Section{
+		{Name: "config.yaml", Content: "redacted", Redacted: true},
+	})
+
+	mock.ExpectQuery(regexp.QuoteMeta(`SELECT instance_id, reason, captured_at, sections`)).
+		WithArgs("ws-1", "org-9").
+		WillReturnRows(sqlmock.NewRows([]string{"instance_id", "reason", "captured_at", "sections"}).
+			AddRow("i-abc", "bootstrap_watcher", ts, secs))
+
+	got, err := NewPostgres(dbh).GetLatest(context.Background(), "ws-1", "org-9")
+	if err != nil {
+		t.Fatalf("GetLatest: %v", err)
+	}
+	if got == nil {
+		t.Fatal("got nil, want a bundle")
+	}
+	if !got.CapturedAt.Equal(ts) {
+		t.Errorf("captured_at = %v, want %v", got.CapturedAt, ts)
+	}
+	if got.Bundle.InstanceID != "i-abc" || got.Bundle.Reason != "bootstrap_watcher" {
+		t.Errorf("bundle meta wrong: %+v", got.Bundle)
+	}
+	if len(got.Bundle.Sections) != 1 || got.Bundle.Sections[0].Name != "config.yaml" {
+		t.Errorf("sections decoded wrong: %+v", got.Bundle.Sections)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet: %v", err)
+	}
+}
+
+// TestGetLatest_NoRowsReturnsNil: no bundle → (nil, nil), so the handler
+// can 404 without treating it as an error.
+func TestGetLatest_NoRowsReturnsNil(t *testing.T) {
+	dbh, mock := newMock(t)
+	mock.ExpectQuery(regexp.QuoteMeta(`SELECT instance_id, reason, captured_at, sections`)).
+		WithArgs("ws-none", "org-9").
+		WillReturnError(sql.ErrNoRows)
+
+	got, err := NewPostgres(dbh).GetLatest(context.Background(), "ws-none", "org-9")
+	if err != nil {
+		t.Fatalf("GetLatest err = %v, want nil for no-rows", err)
+	}
+	if got != nil {
+		t.Fatalf("got %+v, want nil for no-rows", got)
+	}
+}
+
+// TestGetLatest_OrgScopingArg: the org id is passed as the $2 filter arg
+// with strict equality, so a row in a sibling org is excluded by the query
+// itself. A mismatched org → no row → nil (same as no-rows).
+func TestGetLatest_OrgScopingArg(t *testing.T) {
+	dbh, mock := newMock(t)
+	// Tenant org-B asks for ws-1 (owned by org-9). The strict predicate
+	// filters it out → ErrNoRows → nil.
+	mock.ExpectQuery(regexp.QuoteMeta(`AND org_id = $2`)).
+		WithArgs("ws-1", "org-B").
+		WillReturnError(sql.ErrNoRows)
+
+	got, err := NewPostgres(dbh).GetLatest(context.Background(), "ws-1", "org-B")
+	if err != nil {
+		t.Fatalf("GetLatest: %v", err)
+	}
+	if got != nil {
+		t.Fatal("sibling-org read returned a bundle; want nil")
+	}
+}
+
+// TestGetLatest_EmptyOrgIDRejected: an empty orgID must fail closed with
+// an error rather than disabling the org filter (#2020).
+func TestGetLatest_EmptyOrgIDRejected(t *testing.T) {
+	dbh, _ := newMock(t)
+	_, err := NewPostgres(dbh).GetLatest(context.Background(), "ws-1", "")
+	if err == nil {
+		t.Fatal("GetLatest(empty orgID) should error")
+	}
+}
+
+// TestGetLatest_QueryErrorPropagates: a real DB error (not ErrNoRows)
+// surfaces as an error so the handler returns 503, not a false 404.
+func TestGetLatest_QueryErrorPropagates(t *testing.T) {
+	dbh, mock := newMock(t)
+	mock.ExpectQuery(regexp.QuoteMeta(`SELECT instance_id, reason, captured_at, sections`)).
+		WithArgs("ws-1", "org-9").
+		WillReturnError(errors.New("connection reset"))
+
+	_, err := NewPostgres(dbh).GetLatest(context.Background(), "ws-1", "org-9")
+	if err == nil {
+		t.Fatal("want an error for a non-ErrNoRows DB failure")
+	}
+}
+
+// TestNilDB: both methods return an error (never panic) when the db
+// handle is nil — the degraded-boot guard the wiring relies on.
+func TestNilDB(t *testing.T) {
+	p := NewPostgres(nil)
+	if err := p.Persist(context.Background(), sampleBundle()); err == nil {
+		t.Error("Persist(nil db) should error")
+	}
+	if _, err := p.GetLatest(context.Background(), "ws-1", "org-9"); err == nil {
+		t.Error("GetLatest(nil db) should error")
+	}
+}
@@ -703,6 +703,14 @@ func Setup(hub *ws.Hub, broadcaster *events.Broadcaster, prov *provisioner.Provi
 	wsAuth.PUT("/files/*path", tmplh.WriteFile)
 	wsAuth.DELETE("/files/*path", tmplh.DeleteFile)

+	// Rescue read (RFC internal#742 Part 3) — latest post-mortem bundle
+	// for a boot-failed/terminated workspace, so "why won't my agent
+	// boot" is answerable without a live instance. Same WorkspaceAuth
+	// gate as /files/*; the handler org-scopes the store read by
+	// MOLECULE_ORG_ID so a sibling org cannot read another org's bundle.
+	rescueReadH := handlers.NewRescueReadHandler()
+	wsAuth.GET("/rescue", rescueReadH.GetRescue)
+
 	// Chat attachments — file upload (user → agent) and binary-safe
 	// streaming download (agent → user). Namespaced under /chat/ so
 	// the security model is obviously distinct from /files/* (which
@@ -0,0 +1,4 @@
+-- Reverse RFC internal#742 Part 3 rescue_bundles table.
+-- Forensic-only table; dropping it loses post-mortem read history but
+-- does not affect boot-failure semantics (capture still ships to Loki).
+DROP TABLE IF EXISTS rescue_bundles;
--- a/Show More
+++ b/Show More