fix(cat): forget_memories actually deletes + skip guaranteed-413 Groq round-trips

[catomean]

Incident (prod, 2026-08-02, user @mao)

User: "Photography, handmade ceramics, and French don't apply to me, and the weekends-only availability is wrong. Forget all of those."

The model emitted a well-formed forget_memories exec_action with exactly those phrases — twice — the action logged completed (economic-profile entries were removed), yet every targeted cat_memories row survived: "Is a professional photographer who speaks fluent French…", "Has a documentary photography background", "Selling a handmade ceramic coffee mug…", "Knows French", "Can only work on weekends."

Root causes (both verified against prod)

1. Lexical matcher can't see inflections. Exact-substring token matching: photography ≠ photographer, ceramics ≠ ceramic, speaking ≠ speaks — plus a hard ≥2-token-hit rule. All four phrases scored 0–1 hits → no match.
2. Semantic floor was unreachable. FORGET_MATCH_SIMILARITY = 0.75, but embedding the actual incident phrases against the actual stored rows (text-embedding-3-small, prod embeddings) scores true targets 0.39–0.61 and unrelated rows ≤ 0.29. The fallback could never fire for real phrasings.

Fix

• Light suffix-stripping stemmer (photography/photographer → photograph; equality-only stem compare, so constraint can never collide with construction) + majority-of-significant-stems rule (1 of 1, 1 of 2, 2 of 3…).
• Semantic floor 0.75 → 0.45 (sits in the measured gap between targets and noise).
• displayMessage now lists exactly what was removed, not just counts — an over-match is visible and re-addable at a glance.
• Regression suite pins the incident corpus + the exact phrases the model passed, including the over-deletion guard: "weekend availability constraint" must NOT lexically nuke the "extra income on weekends" goal.

Second fix: guaranteed-failing Groq round-trip on every message

journalctl showed every Cat chat failing on primary Groq with HTTP 413 and falling back to OpenRouter. Measured against the platform key: Groq on-demand tier hard-rejects requests over 12,000 TPM — and a grown Cat prompt (22 memories + history + page excerpt) exceeds that deterministically. The chain now pre-flight-skips a platform Groq link that clearly can't fit, only when another link can serve; BYOK Groq is always attempted, and a skip never poisons the global link-health breaker.

Verification

• npm run verify green: typecheck clean, lint 0 errors, 145 suites / 1467 tests pass (13 new).
• Similarity numbers measured against the live prod DB + real OpenAI embeddings; TPM limit measured with a real 413 against the platform key.

🤖 Generated with Claude Code

[catomean]

Closing as superseded — this work is already on main, in a more complete form.

Checked before closing, rather than assuming:

• Groq 413 pre-flight — present in src/services/ai/groq.ts on main, and __tests__/unit/ai/groq-preflight.test.ts exists there too.
• Forget matching (stemmer + lowered similarity floor) — present in src/services/cat/memory.ts on main, with more of it than this branch carries (20 matching references vs 10 here), via fix(cat): forgetting a two-word fact could delete an unrelated memory #579 and the surrounding fixes.
• Every symbol this branch introduces — stemWord, STEM_SUFFIXES, factStems, memoryStems, significantStems, FORGET_MATCH_SIMILARITY, GROQ_ON_DEMAND_TPM_LIMIT, promptFitsGroqOnDemand, overflowsPlatformGroq, MAX_LISTED — already resolves on main. Nothing here is unique.

It also conflicts in src/services/cat/memory.ts and __tests__/unit/cat/memory-forget.test.ts, and because main's version is the newer and larger one, resolving those conflicts would risk reverting improvements rather than adding any.

This is the parallel-session duplicate-work pattern: two sessions fixed the same bug on different branches, and the other one landed first. Closing so it stops blocking the auto-merge queue.