Targeted coverage uplift across the framework core's pure helpers and
two integration boundaries. Local src/runtime coverage rises from
87.21% to 88.90% (+1.69%); 97 previously-uncovered lines now exercised.

New test files (7):

  * tests/test_llm_stub_structured_output.py
      StubChatModel.with_structured_output happy path
      (llm.py:141-160, 171-177). The defensive model_validate fallback
      (lines 161-169) is excluded with a documented justification:
      pydantic v2's model_validate calls __init__, so any schema whose
      constructor raises also fails the fallback.

  * tests/test_envelope_recovery.py
      _try_recover_envelope_from_raw (graph.py:583-610). Table-driven
      across the three candidate-substring strategies (raw, fenced,
      greedy first-{...-last-}) and every failure path.

  * tests/test_orchestrator_extract_last_error.py
      Orchestrator._extract_last_error (orchestrator.py:945-998).
      Pure mapping from a failed-AgentRun summary string to a
      representative typed exception. Table-driven across
      EnvelopeMissingError, ValidationError, TimeoutError,
      OSError, RuntimeError fallback, plus reversed-iteration ordering.

  * tests/test_handle_agent_failure.py
      _handle_agent_failure (graph.py:613-644). Both the happy path
      (reload + append + status='error') and the FileNotFoundError
      fallback path (use caller's in-memory session). Plus a
      partial-tool-write preservation regression test.

  * tests/test_retry_session_locked_post_policy.py
      Orchestrator._retry_session_locked post-policy execution path
      (orchestrator.py:1552-1587). Stub orchestrator pulls in the
      real method body and substitutes the surrounding integration
      points (graph, finalize, pause). Covers the failed-AgentRun
      filter, retry_count + active_thread_id pinning, and the
      pause-vs-finalize fork.

  * tests/test_service_run_exception_branches.py
      OrchestratorService.start_session._run exception branches
      (service.py:541-568). Three classes get distinct treatment:
      CancelledError (propagate as-is), GraphInterrupt (propagate
      WITHOUT marking registry status='error' -- HITL pause is not a
      failure), generic Exception (mark error then propagate).

  * tests/test_sse_tail_loop.py
      SSE _stream tail-poll loop (api.py:879-890), including the
      CancelledError re-raise from PR #13. Two tests: one drives the
      tail to deliver a post-drain event, one forces sleep to raise
      CancelledError and asserts it propagates (pinning the bug fix).

Verification:
  ruff check src/ tests/                         passed
  pyright src/runtime                            0 errors / 0 warnings
  pytest -x                                      1310 passed / 8 skipped (was 1265/8)
  pytest --cov=src/runtime --cov-fail-under=85   88.90% (was 87.21%)
  build_single_file.py                           dist unchanged (tests only)

Projected SonarCloud impact (after the Phase 1 PR #13 exclusion sync):
  new_coverage   ~87.2% -> ~89-90%