fix(autoCombo,sse): drop vendor-retired ids from the auto-combo pool and skip leftover arena_elo (#11625)

[MumuTW]

Fixes #11625

Not a replacement for #11598. That PR's layer-3 veto (prefix / inherited-base) stays on #11508. This is the candidate pool + layers 1–2, which still select a retired model even after a layer-3 veto.

⚠️ base-red inherited: #11449 — remaining red checks on this PR reproduce on the pristine base tip.

Problem

getTaskFitnessWithSource() is user_override → arena_elo → inherited 1–2 → models_dev_tier → FITNESS_TABLE → wildcard.

#11507 stopped layer 4 from ranking retired ids at 0.98. #11508 / #11598 veto layer 3. Neither stops auto-combo from picking a retired model:

1. buildAutoCandidates() only drops dashboard-hidden models. filterSelectableModels() exists but reads MODEL_LIFECYCLE_RECORDS (OpenAI-only hardcoded). An Anthropic/Google retired id is untracked → allow.
2. A leftover arena_elo / user_override row returns at layer 1–2, so the layer-3 veto never runs. lookupInheritedFitness also copies a dead base's arena score onto an effort variant.
3. Aggregator traffic (openrouter + gpt-5.2-codex) is untracked because the dated table is provider-scoped.

allowedRetiredInCatalog (6 ids) stays a catalog/CI burn-down. This PR does not delete catalog rows; it stops auto-combo from selecting them and stops leftover intelligence rows from scoring above the 0.5 neutral.

Fix

One prefix-stripped predicate, isVendorRetiredId(), reads config/quality/model-lifecycle.json (status === "retired" only — retiring with retiredOn: null is left alone).

• rejectRetiredAutoComboCandidates() runs after the hidden-model filter in buildAutoCandidates.
• getTaskFitnessWithSource skips layers 1–3 when the request id or its resolveScoresAs() base is retired, and returns wildcard 0.5 with no *codex / code boost.
• getModelLifecycleDecision consults the same snapshot after the dated OpenAI table misses, so aggregator traffic is not untracked for ids the JSON already knows. Dated OpenAI warn-before-shutdown rows stay provider-scoped.

Tests

TDD (Hard Rule #18): tests/unit/retired-pool-11625.test.ts was written first and failed 9/11 on the pre-fix chain.

• isVendorRetiredId("openai/gpt-5.2-codex") and the bare form
• rejectRetiredAutoComboCandidates keeps gpt-5.6-sol, drops snapshot-retired ids
• isModelSelectable("anthropic", "claude-3-7-sonnet-20250219") === false
• seeded arena_elo on gpt-5.2-codex / openai/gpt-5.2-codex is not returned
• claude-3-7-sonnet-20250219-high does not inherit the dead base's arena row
• live gpt-5.6-sol arena row still wins

Existing model-lifecycle.test.ts updated: dated OpenAI deprecations stay provider-scoped; snapshot-retired ids are rejected on every provider. #11489 inherit tests still pass.

Verification

DISABLE_SQLITE_AUTO_BACKUP=true node --import tsx/esm --import ./open-sse/utils/setupPolyfill.ts --import ./tests/_setup/isolateDataDir.ts --test \
  tests/unit/retired-pool-11625.test.ts tests/unit/model-lifecycle.test.ts tests/unit/task-fitness-scores-as-11489.test.ts
# tests 22 | pass 22 | fail 0

DISABLE_SQLITE_AUTO_BACKUP=true node --import tsx/esm --import ./open-sse/utils/setupPolyfill.ts --import ./tests/_setup/isolateDataDir.ts --test \
  tests/unit/model-lifecycle-integration.test.ts tests/unit/auto-combo-hidden-models-4558.test.ts tests/unit/model-deprecation-aliases-11503.test.ts
# tests 73 | pass 73 | fail 0

npm run check:model-lifecycle
# PASS — snapshot 2026-08-25, 68 retired id(s), 1327 catalog id(s)

npm run check:cycles
# OK - no cycles detected across 429 files

npm run typecheck:core reports the pre-existing omniglyph package-export mismatches on files this PR does not touch (same baseline as #11492 / #11507). Touched files are under the 1000-line cap.