Bug: silent JSON-parse failures drop triple extractions permanently

[raphasouthall]

Summary

Triple extraction silently fails on malformed LLM JSON output with no retry. Once a note fails, it stays at zero triples forever unless the user manually runs backfill. Audit of the vault shows 25 notes (6%) with no triples — 5 of these are substantive content (>5 KB, one 14 KB).

Root cause

Two layers of try/except that swallow the failure, plus a write-skip on empty triples with no sentinel row:

src/neurostack/triples.py:86-90:
```python
try: triples = json.loads(raw)
except json.JSONDecodeError as e:
log.warning(...)
return []
```

src/neurostack/watcher.py:475-488 wraps the call in a bare try/except that also swallows. Then watcher.py:578 only writes triples if triples: — empty list = nothing inserted, no sentinel row, no retry scheduled.

Result: indistinguishable in the DB from notes that legitimately produce no triples (e.g., 42-byte stubs).

Proposed fix

1. JSON mode + one-shot retry. In triples.py:65-77, add "response_format": {"type": "json_object"} (supported by Ollama and OpenAI-compatible endpoints). On parse failure, retry once with a follow-up turn.
2. Sentinel / retry queue. In watcher.py:_index_triples_for_note (line 369) and _write_note_results (line 578), when extract_triples returns [] on content >200 chars, write a triples_failed row (or a boolean column on notes) with next_retry_at. neurostack backfill picks these up with exponential backoff.

Expected effect

Triple coverage climbs from 94% → 99%+ without user intervention. Silent failures become observable via a retry queue.

Key files

• src/neurostack/triples.py:65-90
• src/neurostack/watcher.py:369, 475-488, 578, 848-905

[raphasouthall]

Acceptance criteria

• vault_stats triple_coverage ≥ 99% on a freshly re-indexed vault
• New `triples_failed` table (or triples_pending column on `notes`) with `next_retry_at`
• Integration test: inject malformed LLM JSON → assert retry row written, no exception bubbled
• Integration test: subsequent valid response clears the retry row and persists triples
• neurostack doctor or neurostack backfill surfaces pending-retry count
• triples.py uses response_format: json_object where the backend supports it, with one-shot retry on parse failure

Dependencies

• Blocks: improves evidence quality for Bug: community hierarchy is inverted (fine partition has fewer clusters than coarse) #33 (communities use triples via co-occurrence) and Improvement: co-occurrence table is 62% of DB and 82% noise #34 (co-occurrence entity graph builds from triples).
• Depends on: none. Independent bug fix.

Sequencing

Phase 1 (quick wins). Fix early — downstream signal cleanup is only trustworthy once triple extraction is reliable.