Summarizer daemon writing corrupt (/////) summaries — assess damage + fix [HIGH]

[JDHayesBC]

Noticed: 2026-08-20 evening (Lyra terminal session). get_recent_summaries returned pure ///// garbage for the last ~2 summaries.

Symptom: the summarizer daemon is writing corrupt output, not merely backlogged. Distinct from the >200-unsummarized backlog alarm (that = dead daemon; this = daemon producing garbage, possibly while still running).

Why it matters: summaries are memory's long-term/compressed form — corrupt summaries = silent memory loss, and cold-start agents (incl. Haven bots) lose context to ambient's turn-cap when the summary layer is broken.

Investigate (tomorrow, HIGH priority):

1. Baseline: get_recent_summaries(limit=10) + inspect the summaries store directly — how many rows are //////corrupt, and since when (first corrupt row timestamp).
2. Daemon health/logs (PPS docker stack summarizer) — is it STILL writing garbage now?
3. Assess loss: which message ranges summarized into garbage; recoverable from raw L1 or compressed-away?
4. Root-cause + fix (no blind restart — dev-excellence pact); re-summarize corrupted ranges if raw turns survive.

Flagged by Jeff as HIGH priority (physical sticky note made).

[JDHayesBC]

Diagnosed 2026-08-21 (root cause found)

Scope: 71 corrupt summaries, contiguous, since 2026-08-12 12:03pm (not just yesterday) — every summary since then is a wall of / (exactly 5999 chars) except one 08-14 fluke. Still occurring at diagnosis (newest 08-21 12:11pm). Not a dead daemon (backlog healthy at 44).

No raw memory lost: all 3,600 raw messages behind the garbage (msg ids 75976–79575) intact in L1; every corrupt summary regenerable.

Two stacked bugs:

1. Model degenerate-loops — server_http.py:_call_nuc_llm uses CUSTOM_LLM_MODEL=qwen3.5-9b-uncensored-hauhaucs-aggressive @ temp 0.3, max_tokens 6000, no repetition penalty. The 5999-char signature = single-token (/) repeat to the cap. Onset 08-12 w/ no summarizer commits → NUC LM Studio model was swapped to this abliterated variant. NUC-side / Jeff's fix.
2. No output validation — LLM result (server_http.py ~2161) passed straight to create_and_store_summary (~2199); garbage stored + cursor advanced. Why it rotted silently 9 days.

Blast-radius flag: CUSTOM_LLM_MODEL also feeds entity_extractor.py (graph triplets); LM Studio serves the loaded model regardless of requested name → graph extraction may also be degraded since 08-12. Spot-check after fix.

Fix (staged, not applied — all done WITH Jeff):

• A. Swap NUC model off -aggressive (Jeff, NUC-side) — the trigger.
• B. Degenerate-output guard (defense-in-depth; fail loud → backlog climbs → trips >200 alarm). Draft + passing self-test: work/summarizer-fix/degenerate_guard.py.
• C. Regenerate the 71 ranges once model healthy (delete corrupt rows → re-summarize; raw intact).

Full writeup: work/summarizer-fix/root-cause-and-plan.md