fix(autoCombo,sse): catalog hygiene — retire dead FITNESS_TABLE rows, fix 7 BUILT_IN_ALIASES targets, add model-lifecycle gate (#11503)

[MumuTW]

Fixes #11503

Two hand-maintained tables decide routing and both had rotted against the vendors' own deprecation pages. This PR cleans them and adds an offline gate so the next drift is a red check instead of a silent ranking inversion.

---

Commit 1 — FITNESS_TABLE hygiene + segment-boundary matching

Problem. Layer 4 of the task-fitness chain was a table of version-less family patterns (codex, claude-sonnet, qwen, llama, gemini-pro, …) matched with String.includes. Recognition by the table, not availability, decided rank: openai/gpt-5.2-codex (OpenAI shut the codex family down 2026-07-23) scored 0.98 for coding via the codex row, while live flagships the table had never heard of (claude-fable-5-thinking-max, gpt-5.6-sol-xhigh) sat at the wildcard 0.50. Substring matching leaked further: o3 matched solar-pro3, and gpt-4o matched chatgpt-4o-latest — a different, retired model that inherited the live flagship's 0.9.

Fix.

• Dropped every row that matches nothing in the catalog or names a retired family (o1, mixtral, grok-4-fast, codex) and every version-less family row (claude-sonnet, claude-opus, claude-haiku, qwen, llama, mistral, gemini-pro, gemini-flash, grok-4, kimi-k2, glm-5). Versioned rows naming a live model stay (gpt-4o, gpt-4o-mini, gpt-4-turbo, o3, o4-mini, gemini-2.5-*, gemini-3.1-pro, deepseek-*, grok-3, glm-5.1, minimax-*).
• getStaticFitnessTableScore now matches on segment boundaries (- / . / / / start / end), longest pattern first as before. o3 still matches o3, o3-mini, openai/o3 but no longer solar-pro3; gpt-4o still matches gpt-4o-2024-05-13 but no longer chatgpt-4o-latest — that one now falls to 0.5, which is the intended outcome since OpenAI shut it down 2026-02-17.
• The file header now documents that layer 4 is a small versioned table and that the wildcard 0.5 an unknown id falls to means "no evidence", never a quality claim. scoring.ts is untouched — the fix is removal, not reweighting.

Tests. New open-sse/services/autoCombo/__tests__/fitness-table-hygiene-11503.test.ts (vitest): loads the lifecycle snapshot and asserts no routable retired id scores through the table at all; solar-pro3, chatgpt-4o-latest, claude-sonnet-5, claude-fable-5-thinking-max, gpt-5.6-sol-xhigh → null; gpt-4o-mini still 0.8, o3-mini still 0.95 via o3, openai/o3 still resolves.

Before / after — every catalog id resolved through getTaskFitnessWithSource(id, "coding")

	before (a179ffed5b)	after
resolved via fitness_table	533 / 1327	73 / 1327
resolved via wildcard_boost	794	1254
retired ids scoring ≥ 0.88 via fitness_table	2 (openai/gpt-5.2-codex 0.98, chatgpt-4o-latest 0.90)	0

⚠️ The issue's audit quoted larger counts (12 ids ≥ 0.88). Those figures folded in ids the vendors mark retiring (shutdown scheduled) alongside retired, and counted bare ids the catalog does not actually route. Counting only ids that are both status: "retired" and present in REGISTRY, the reproducible numbers on this base are 8 routable retired ids, 2 of them ≥ 0.88. Both are now 0.

Commit 2 — BUILT_IN_ALIASES targets + provider guard

Problem. The alias table rewrites body.model on every request (modelLifecyclePolicy.ts), so a stale target is a guaranteed 404. Seven rows pointed at a retired or non-existent model — five Claude rows chained one retired id to another.

source	old target	status	new target
claude-3-opus-20240229	claude-opus-4-20250514	retired 2026-06-15, absent	claude-opus-4-8
claude-3-sonnet-20240229	claude-sonnet-4-20250514	retired 2026-06-15, absent	claude-sonnet-4-6
claude-3-5-sonnet-latest	claude-sonnet-4-20250514	retired 2026-06-15, absent	claude-sonnet-4-6
claude-3-haiku-20240307	claude-3-5-sonnet-20241022	retired 2025-10-28	claude-haiku-4-5-20251001
claude-3-5-haiku-latest	claude-3-5-sonnet-20241022	retired 2025-10-28	claude-haiku-4-5-20251001
gemini-3-pro-high	gemini-3.1-pro-high	not a catalog id	gemini-3-1-pro-high (the catalog's spelling)
llama-3-8b	llama3-8b-8192	Groq deprecated 2025-08-30, absent	llama-3.1-8b-instant

Every replacement is the one the vendor publishes; the other 23 rows are untouched.

Provider guard. The table is global but the catalog is not — aggregators still serve kimi-k2, gemini-2.0-flash, mistral-large under their original ids, and the unconditional rewrite broke them. resolveModelAlias(modelId, provider?) now returns modelId unchanged when the given provider serves it as-is (hasKnownProviderModel, exported from open-sse/services/model.ts for this). Both modelLifecyclePolicy.ts call sites pass the provider they already have. getDeprecationNotice / isDeprecated keep their 1-arg behaviour, and a custom (operator-authored) alias still wins over a built-in one, provider or not.

Tests. New tests/unit/model-deprecation-aliases-11503.test.ts (node:test, 66 assertions): table-driven over every BUILT_IN_ALIASES entry — each target must be a catalog id (checked against REGISTRY, not a hardcoded list) and must not be retired in the snapshot; plus resolveModelAlias("kimi-k2", "t3-web") === "kimi-k2", resolveModelAlias("kimi-k2", "fireworks") === "moonshotai/Kimi-K2", resolveModelAlias("claude-3-opus-20240229") === "claude-opus-4-8", and custom-alias precedence.

Commit 3 — check:model-lifecycle gate + vendor snapshot

config/quality/model-lifecycle.json — 106 lifecycle entries (68 retired) curated from the five first-party deprecation pages listed in its sources. replacement is null wherever the vendor publishes none.

scripts/check/check-model-lifecycle.mjs — no network, three checks, all summed before exit:

1. no FITNESS_TABLE pattern scores a routable retired id, under the new boundary rule;
2. no BUILT_IN_ALIASES target is retired or absent from the catalog;
3. every retired id the catalog still routes is either forwarded by BUILT_IN_ALIASES or listed in allowedRetiredInCatalog.

Check (1) is deliberately scoped to routable ids: legitimate versioned rows like gpt-4o also match retired ids the catalog never served (gpt-4o-audio-preview), and those cannot invert any routing decision. The gate sets DATA_DIR to a throwaway temp dir before importing production modules so it can never migrate the operator's database.

scripts/quality/refresh-model-lifecycle.mjs — network refresher, not wired into CI. It parses Anthropic's markdown table (| API model name | Current state | Deprecated | Tentative retirement date |); for the other vendors it keeps the existing entries and prints manual review needed rather than guessing at a shifting HTML layout. It aborts rather than write a snapshot with fewer entries than it started with — it must never silently drop one.

Wired into .github/workflows/quality.yml (gates=(…)) and .github/workflows/ci.yml's lint job — the latter is beyond the letter of the issue but is where the other pass/fail policy gates live, so a PR to main gets the same protection. Documented in docs/architecture/QUALITY_GATES.md. New scripts: check:model-lifecycle, quality:refresh-model-lifecycle.

Tests. tests/unit/check-model-lifecycle-gate.test.ts exercises each of the three checks plus the catalog/retired helpers against small fixtures.

allowedRetiredInCatalog — the burn-down list

Check (3) is red on the current catalog, which is the point. To keep this PR mergeable the 6 offenders are allowlisted with a TODO(#11503) note in the file; the maintainer's call is remove-from-catalog vs. add-a-forward for each:

id	vendor status
chatgpt-4o-latest	OpenAI, retired
claude-3-5-sonnet-20241022	Anthropic, retired 2025-10-28
claude-3-7-sonnet-20250219	Anthropic, retired 2026-02-19
google/gemini-2.0-flash	Google, retired 2026-06-01 (the bare gemini-2.0-flash is forwarded; the prefixed form is not, since the alias table is exact-id keyed)
gpt-4-0125-preview	OpenAI, retired
openai/gpt-5.2-codex	OpenAI, retired

Adjacent but deliberately not changed here: llama-3.1-8b-instant (the new llama-3-8b target) is marked retiring by Groq with a 2026-08-16 date. It is the replacement Groq itself names and is live in the catalog, so it is the correct target today; the gate will flag it the moment the page flips it to retired.

---

Deviations from the issue's proposed scope

• The lifecycle snapshot lands in commit 1, not commit 3, because commit 1's test consumes it — this keeps every commit green in isolation.
• Existing suites that used the removed claude-sonnet / claude-opus rows as their "known static model" probe were retargeted to a surviving versioned row (o3 carries exactly the same coding 0.95 / review 0.92 numbers; gemini-2.5-pro covers one review case). The behaviour under test — layer-4 resolution, longest-pattern-first, default-table fallback, case-insensitivity — is unchanged, and the fix(backend): static fitness table substring match returns wrong row for longer model ids (gpt-4o-mini scores as gpt-4o) #8603 regression assertions are intact.
• tests/unit/model-deprecation.test.ts had pinned the two stale Claude targets; updated to the corrected ones.

Verification

command	result
npx vitest run --config vitest.mcp.config.ts open-sse/services/autoCombo/__tests__/	Test Files 7 passed (7) · Tests 106 passed (106)
node --test tests/unit/{task-fitness-missing-table-8603,model-deprecation-aliases-11503,check-model-lifecycle-gate}.test.ts	# tests 77 · # pass 77 · # fail 0
node --test tests/unit/{model-deprecation,model-alias-seed,model-combo-mappings,model-lifecycle,model-lifecycle-integration,8676-monsterapi-deprecation,gemini-cli-deprecation}.test.ts	119 pass · 0 fail
npm run check:model-lifecycle	PASS — snapshot 2026-08-25, 68 retired id(s), 1327 catalog id(s) (without the allowlist: 6 violation(s), listed above)
npm run check:cycles	OK - no cycles detected across 429 files
npm run typecheck:core	clean
npx eslint <touched files> --suppressions-location config/quality/eslint-suppressions.json	0 errors
npm run check:known-symbols / check:docs-sync / check:fabricated-docs / check:test-discovery	all PASS
npm run check:docs-counts	✗ 3 stale "159 migrations" refs (README/AGENTS/llm.txt) — pre-existing on the base, unrelated

⚠️ base-red inherited: #11449 (the base tip is red for unrelated reasons; #11502 clears most).