feat(M91): Smart Retry for Divergent Samples

[hlin99]

Summary

Adds smart-retry capability that automatically reruns divergent samples with deterministic settings (temperature=0, seed=42) to classify each divergence as deterministic (real bug) or stochastic (sampling noise).

Changes

• smart_retry.py: SmartRetryResult, SampleRetryResult, run_smart_retry(), format_smart_retry()
• CLI: xpyd-acc smart-retry --report --baseline --target [--json]
• Exit 1 if deterministic divergences found (CI-friendly)
• ROADMAP.md: marked M87/M89 complete, added M91
• 12 tests covering retry logic, classification, formatting, JSON export

Closes #194

[hlin99-Review-Bot]

Review: Request Changes

Idea: ✅ Smart retry to classify divergences as deterministic vs stochastic is a solid addition — reduces false positives in CI and gives actionable signal.

Code issue — CI failing (6 test failures on Python 3.11):

Tests use asyncio.get_event_loop().run_until_complete() which raises RuntimeError: There is no current event loop in thread 'MainThread' on Python ≥3.10 (deprecated) / 3.12+ (removed).

Fix: Replace all instances in tests/test_smart_retry.py:

# Before
result = asyncio.get_event_loop().run_until_complete(
    run_smart_retry(report, ...)
)

# After
result = asyncio.run(
    run_smart_retry(report, ...)
)

This affects tests at lines 128, 136, 157, 181, 203, 225.

Everything else looks clean — module structure, dataclasses, CLI integration, formatting all LGTM. Just fix the async test pattern and CI should go green.

[hlin99-Review-BotX]

Idea: ✅ Good — Smart retry to classify divergences as deterministic vs stochastic is valuable. Aligns well with the project's accuracy tooling goals.

Code: Needs fix — CI is red. All 6 async tests in tests/test_smart_retry.py fail on Python 3.11 with:

RuntimeError: There is no current event loop in thread 'MainThread'.

The tests use asyncio.get_event_loop().run_until_complete(), which was deprecated in 3.10 and raises on 3.11+. Replace with asyncio.run():

# Before
result = asyncio.get_event_loop().run_until_complete(
    run_smart_retry(report, "http://base", "http://target")
)

# After
result = asyncio.run(
    run_smart_retry(report, "http://base", "http://target")
)

Fix the 6 occurrences, confirm CI passes, then re-request review.

[hlin99-Review-Bot]

CI green across all Python versions (3.10/3.11/3.12). The asyncio fix looks correct — asyncio.run() is the right replacement. Code, tests, and docs all good. ✅ Approved.

[hlin99-Review-BotX]

CI green across all Python versions. The asyncio.run() fix is correct — addresses the exact issue I flagged. Code, tests, docs all look good. ✅ Approved.