feat(tests): GAP-05 add _get_with_retry() with 429 back-off + fix broken test_call_peer_errors

Adds retry-on-429 with exponential back-off (1 s → 2 s → 4 s, ±25% jitter,
30 s cap, Retry-After header honoured) to all idempotent RemoteAgentClient
GET calls: poll_state, pull_secrets, get_peers, discover_peer.

Also fixes the merged test_call_peer_errors.py (PR #7) which was broken:
- Removed pytest-mock dependency (mocker not installed)
- Fixed call_peer(message: str) vs dict
- Fixed non-existent _call_direct/_call_proxy method patches
- Uses FakeResponse + _session.post.side_effect pattern consistently

Adds tests/conftest.py (FakeResponse + client fixture + _CaptureHandler)
and tests/test_retry_backoff.py (18 new tests).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/conftest.py

@@ -1,6 +1,6 @@
 """Pytest fixtures and helpers for molecule_agent tests.
 
-All fixtures are pytest-scoped unless noted.  No live platform required —
+All fixtures are function-scoped unless noted.  No live platform required —
 all HTTP is mocked via ``unittest.mock``.
 """
 from __future__ import annotations
@@ -28,10 +28,12 @@ class FakeResponse:
         status_code: int = 200,
         json_body: Any = None,
         text: str = "",
+        headers: dict[str, str] | None = None,
     ) -> None:
         self.status_code = status_code
         self._json = json_body
         self.text = text
+        self.headers = headers or {}
 
     def json(self) -> Any:
         return self._json
@@ -104,7 +106,7 @@ class _CaptureHandler:
 
     @classmethod
     def handle(cls, method: str, url: str, **kwargs: Any) -> FakeResponse:
-        for m, p, status, headers, body in reversed(cls._stubs):
+        for m, p, status, hdrs, body in reversed(cls._stubs):
             if m == method and p in url:
-                return FakeResponse(status, json_body={}, text=body)
-        raise RuntimeError(f"no stub for {method} {url}")
+                return FakeResponse(status, json_body={}, text=body, headers=hdrs)
+        raise RuntimeError(f"no stub for {method} {url}")

tests/test_call_peer_errors.py

@@ -1,25 +1,79 @@
-"""GAP-03: call_peer error paths — documents and tests the error surface.
+"""GAP-03 / GAP-11: call_peer error paths — documents and tests the error surface.
 
-Per PLAN.md backlog #13: ClaudeSDKExecutor surfaces opaque "Command failed"
-without capturing stderr. These tests document the desired behavior for the
-SDK's call_peer method in molecule_agent/client.py.
-
-The tests use the ``client`` fixture (MagicMock session) to simulate error
-conditions without a live platform.
+Per PLAN.md backlog #13: call_peer must surface structured errors (HTTP
+status, auth context) rather than opaque strings. These tests verify the
+error surface using the same FakeResponse / MagicMock pattern as the rest of
+the test suite.
 """
 from __future__ import annotations
 
-import sys
+import time
 from pathlib import Path
+from typing import Any
+from unittest.mock import MagicMock
 
 import pytest
 
-_SDK_ROOT = Path(__file__).resolve().parents[1]
-if str(_SDK_ROOT) not in sys.path:
-    sys.path.insert(0, str(_SDK_ROOT))
+from molecule_agent import RemoteAgentClient
 
-from molecule_agent.client import RemoteAgentClient
-from tests.conftest import FakeResponse
+
+# ---------------------------------------------------------------------------
+# FakeResponse — minimal requests.Response stand-in
+# ---------------------------------------------------------------------------
+
+
+class FakeResponse:
+    """Minimal stand-in for ``requests.Response``."""
+
+    def __init__(
+        self,
+        status_code: int = 200,
+        json_body: Any = None,
+        text: str = "",
+        headers: dict[str, str] | None = None,
+    ) -> None:
+        self.status_code = status_code
+        self._json = json_body
+        self.text = text
+        self.headers = headers or {}
+
+    def json(self) -> Any:
+        return self._json
+
+    def raise_for_status(self) -> None:
+        if self.status_code >= 400:
+            import requests
+            raise requests.HTTPError(f"HTTP {self.status_code}")
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def tmp_token_dir(tmp_path: Path) -> Path:
+    return tmp_path / "molecule-token-cache"
+
+
+@pytest.fixture
+def client(tmp_token_dir: Path) -> RemoteAgentClient:
+    session = MagicMock()
+    return RemoteAgentClient(
+        workspace_id="ws-test-123",
+        platform_url="http://platform.test",
+        agent_card={"name": "test-agent"},
+        token_dir=tmp_token_dir,
+        session=session,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Error surface tests
+# ---------------------------------------------------------------------------
+
+# Note: call_peer(message: str) — the public API accepts a plain string.
+# Internal A2A envelope is built by the client. Tests pass strings.
 
 
 class TestCallPeerErrors:
@@ -123,7 +177,6 @@ class TestCallPeerErrors:
         - Proxy POST succeeds → result returned
         """
         # Seed the cache so discover_peer returns a URL (cache hit, no GET needed)
-        import time
         client._url_cache["peer-id"] = ("http://dead.peer:8000", time.time() + 60)
 
         post_calls = []
@@ -181,4 +234,4 @@ class TestCallPeerErrors:
         assert "messageId" in body["params"]["message"]
         assert body["params"]["message"]["role"] == "user"
         assert body["params"]["message"]["parts"][0]["kind"] == "text"
-        assert body["params"]["message"]["parts"][0]["text"] == "hello world"
+        assert body["params"]["message"]["parts"][0]["text"] == "hello world"