fix: return best attempt on heal loop exhaustion (v1.14.4)

[devonakelley]

Problem

When the heal loop exhausts all paths, it raises RuntimeError, causing benchmarks to record a hard failure with model="error". A partial or low-quality response is better than an exception for most callers.

Fix

In kalibr/router.py, when heal_result['success'] is False but heal_result['response'] is not None, return the last attempted response instead of raising.

New response attributes set on exhausted return:

• kalibr_healed = True — the heal loop was triggered
• kalibr_heal_exhausted = True — all paths were tried without passing Gate 1/2
• kalibr_heal_count — number of repair attempts made
• kalibr_models_tried — list of models attempted
• kalibr_model_used — final model tried

RuntimeError is still raised if last_response is None (no bytes were ever received — e.g. pure network failure).

Testing

Callers can detect exhaustion via:

response = router.completion(messages=[...], healing=True)
if getattr(response, 'kalibr_heal_exhausted', False):
    # degraded path — log, alert, or fallback
    ...

Checklist

• Fix applied in kalibr/router.py
• Version bumped: 1.14.3 → 1.14.4 in pyproject.toml
• CHANGELOG.md updated
• Do NOT merge without review
• Do NOT publish to PyPI