fix(session): propagate extractor failures to async task error

[dr3243636-ops]

Background

In async session commit (wait=false), extractor exceptions were swallowed by MemoryExtractor.extract() and converted to an empty list.

This caused task tracking to incorrectly report:

• status=completed
• memories_extracted=0
• error=null

which made downstream monitoring (e.g. Slack alerting via task polling) unable to detect real extraction failures.

Changes

1. Added MemoryExtractor.extract_strict()
   • same extraction logic as extract()
   • raises RuntimeError("memory_extraction_failed: ...") on extraction errors
2. Extended SessionCompressor.extract_long_term_memories(..., strict_extract_errors=False)
   • when enabled, it calls extract_strict()
3. Enabled strict mode in async commit path
   • Session.commit_async() now passes strict_extract_errors=True
4. Added test coverage
   • tests/test_session_task_tracking.py::test_task_failed_when_memory_extraction_raises
   • verifies task transitions to failed and exposes error message

Validation

Unit/integration tests

Ran:

python -m pytest tests/test_task_tracker.py tests/test_session_task_tracking.py tests/test_session_async_commit.py -v

Result: 40 passed

Local production verification

Verified on local running OV service before opening this PR:

• Normal model config -> async task completed
• Forced invalid model (claude-haiku-4-5-20251001) -> async task failed with error containing memory_extraction_failed

This confirms failure signals now propagate to task API as expected.

Compatibility

• Existing extract() behavior is unchanged (still non-throwing).
• Strict error propagation is only enabled in async commit path.

[CLAassistant]

All committers have signed the CLA.

[qin-ctx]

[Bug] extract_strict() 是 extract() 的完整复制粘贴（~115 行），仅在 except 块中有行为差异（raise RuntimeError vs return []）。

这意味着后续对 extract() 的任何 bug 修复或功能变更都必须同步到 extract_strict()，否则两者会产生行为分歧（silent divergence）。

建议通过在 extract() 中添加一个 strict 参数来消除重复：

async def extract(
    self, context, user, session_id, *, strict: bool = False
) -> List[CandidateMemory]:
    # ... existing logic ...
    except Exception as e:
        logger.error(f"Memory extraction failed: {e}")
        if strict:
            raise RuntimeError(f"memory_extraction_failed: {e}") from e
        return []

然后 extract_strict 可以只做一层委托，或直接在 compressor 中传 strict=True。

[qin-ctx]

[Suggestion] strict_extract_errors=True 时异常会直接从 extract_long_term_memories 抛出，调用方 commit_async()（session.py:338）没有 try/catch，而是依赖外层 task tracker 来捕获。

这个假设本身是合理的，但建议至少在此处或 commit_async 中加一行注释说明 "extraction errors intentionally propagate to caller (task tracker)"，避免后续维护者误以为遗漏了异常处理而加上 try/catch 吞掉错误。

[dr3243636-ops]

Addressed review feedback in 39c036d:

1. Removed duplicated extraction logic by adding strict flag to MemoryExtractor.extract(...) and making extract_strict(...) a thin compatibility wrapper.
2. Added explicit comment in SessionCompressor.extract_long_term_memories clarifying that strict extraction errors are intentionally propagated to task tracker.
3. Ran formatter + regression tests: tests/test_task_tracker.py, tests/test_session_task_tracking.py, tests/test_session_async_commit.py (40 passed).

Thanks for the review.