[MumuTW]

Fixes #11489

⚠️ base-red inherited: #11449 — every red check on this PR reproduces on the pristine base tip 6435f618f4 with this branch's changes absent (each verified in a detached worktree or by git show base:):

• Docs Gates: README.md / AGENTS.md / llm.txt say "159 migrations", code has 160
• No new ESLint warnings: run-eslint-json.mjs exits 2 on stale entries in config/quality/eslint-suppressions.json (none in files this PR touches)
• Vitest (fast-path): glmCodingProviderConfig.test.ts — 2 failures (glm-5.3-max tiers, 15-vs-14 inventory)
• Unit Tests fast-path 1–4: failures across A2A agent card, provider-nodes, APIKEY_PROVIDERS count, openapi coverage floor, DB handles — none in taskFitness / scoresAs; the only unit test this PR adds passes
• Fast Quality Gates → step "Quality gates (all)": check:lockfile (libxmljs2/…/brace-expansion resolved from registry.npmmirror.com in the base lockfile), check:mutation-test-coverage (2 base test files missing from stryker.conf.json), check:open-sse-typecheck (src/app/api/v1/models/catalog.ts TS2367, baseline 0 → live 2). All other 20 gates pass locally; file-size is base-relative and every touched file is under the 1000-line cap
• Build (advisory): resolveLiveWsUrl / sanitizeLiveWsPort missing exports — untouched by this PR

Problem

Auto-combo task fitness scores every catalog id by exact string match, but the gateway already resolves <model>-<effort> ids to a base model at dispatch. The scoring layer never sees that resolution, so a variant like gpt-5.6-sol-xhigh misses every DB layer and lands on the wildcard 0.5 — while its own base model is scored properly.

Mechanism

getTaskFitnessWithSource() layers 1–2 call getModelIntelligenceBySource(), which is WHERE model = ?. Combo candidate scoring passes the raw registry id (buildAutoCandidates() → parseModel(target.modelStr).model); no effort splitter is imported anywhere under open-sse/services/combo*. Every DB-backed source (arena, models.dev, user overrides) publishes scores for base models only, so no data source can close this gap.

Fix

New open-sse/services/autoCombo/scoresAs.ts exports one resolver:

resolveScoresAs(modelId): { base: string; via: "explicit" | "effort-suffix" | "free-suffix" | null }

Three tiers, in order, and it never guesses:

1. explicit — the registry entry declares the new RegistryModel.scoresAs. One hop only; the target must itself be a catalog id.
2. effort-suffix — a trailing reasoning-effort token stripped by an existing dispatch splitter (splitClaudeEffortSuffix, splitCodexReasoningSuffix). No new regex is introduced.
3. free-suffix — a trailing -free tier marker.

Tiers 2–3 are catalog-anchored: a stripped base is accepted only when findRegistryModelById(base) (newly exported from providerModels.ts — getGlobalModel's exact-id steps 1–2, deliberately not its startsWith step 3) finds it. On today's catalog, 57 of the 201 strippable effort ids strip to a base that does not exist — qwen3.7-max → qwen3.7 (-max is the model name here), grok-4.6-fast-high → grok-4.6-fast, extra-high → extra. The anchor rejects all of them.

getTaskFitnessWithSource() retries user_override then arena_elo against the resolved base on a literal miss and reports <source>:inherited. The score is inherited verbatim — the 12-factor scoring already prices cost and latency per variant, so there is no basis for inventing an effort delta.

The former -free arena-elo special case (lookupFreeAliasArenaElo, source arena_elo_free_alias, #4517) is deleted and folded into tier 3, which also extends it to user_override (it was arena-elo-only before).

Cache keys are the queried model id, so an inherited hit for a variant cannot poison the base's entry or vice versa; invalidateFitnessCache() still clears everything.

FITNESS_TABLE contents are untouched (that is #8777).

Measured effect

Resolving every catalog id + alias (1327 ids) through getTaskFitnessWithSource(id, "coding"), DATA_DIR isolated:

Structural tier breakdown of resolveScoresAs:

tier	ids
explicit	16
effort-suffix	144
free-suffix	7
unresolved	1160

Coverage, with arena_elo seeded on each of the 47 distinct base ids the variants point at (the mechanism's ceiling given full base coverage):

source	before	after
arena_elo	47	47
arena_elo_free_alias	7	— (folded in)
arena_elo:inherited	—	167
fitness_table	511	456
wildcard_boost	762	657

105 ids leave the wildcard entirely; a further 55 move off the static fitness_table onto the real benchmarked signal of the model they actually are. This number rises with every base row any data source adds, which is the point.

scoresAs entries populated

Only relations tiers 2–3 cannot express — 15 entries:

id	scoresAs	justification
gpt-5.6 (openai)	gpt-5.6-sol	Per OpenAI's model reference gpt-5.6 is an alias of gpt-5.6-sol, not a model. Points forward; no suffix stripper can produce this. Siblings -terra/-luna deliberately left alone.
claude-4.6-opus-high (cursor/agy)	claude-opus-4-6	cursor/agy spell Claude ids <version>-<family>; strips to the ghost claude-4.6-opus.
claude-4.6-opus-max	claude-opus-4-6	same
claude-4.6-opus-high-thinking	claude-opus-4-6	same
claude-4.6-opus-max-thinking	claude-opus-4-6	same
claude-4.6-opus-high-thinking-fast	claude-opus-4-6	same
claude-4.6-opus-max-thinking-fast	claude-opus-4-6	same
claude-4.5-opus-high	claude-opus-4-5	same
claude-4.5-opus-high-thinking	claude-opus-4-5	same
claude-4.6-sonnet-medium	claude-sonnet-4-6	same
claude-4.6-sonnet-medium-thinking	claude-sonnet-4-6	same
claude-4.5-sonnet	claude-sonnet-4-5	same (no effort suffix to strip at all)
claude-4.5-sonnet-thinking	claude-sonnet-4-5	same
claude-4-sonnet	claude-sonnet-4	same
claude-4-sonnet-thinking	claude-sonnet-4	same

Everything else stays unresolved by design.

Tests

TDD per Hard Rule #18 — both files were written first and observed failing (Cannot find module '../scoresAs' / 2 of 4 subtests failing) before the implementation landed.

• open-sse/services/autoCombo/__tests__/scoresAs-11489.test.ts (vitest, 10 tests, no DB) — pure resolver. Covers gpt-5.6-sol-xhigh → effort-suffix, gpt-5.6 → explicit, gpt-5.6-luna unresolved (and specifically not gpt-5.6/gpt-5.6-sol), qwen3.7-max unresolved, claude-sonnet-5 never collapsing to its family, the cursor spellings, mimo-v2.5-free → free-suffix, ox-alpha-free unresolved. grok-4.6-fast-high is asserted against findRegistryModelById rather than a hardcoded expectation, so the test stays true as the catalog moves.
• tests/unit/task-fitness-scores-as-11489.test.ts (node:test, 4 tests, DB-backed) — seeds a user_override on gpt-5.6-sol, asserts gpt-5.6-sol-xhigh and gpt-5.6 return it as user_override:inherited and that gpt-5.6-luna does not. Cleans up and calls resetDbInstance() in after().
• open-sse/services/autoCombo/__tests__/autoCombo.test.ts — the feat: qualify ambiguous model names with provider prefix in /v1/models #4517 -free case updated for the new source string (arena_elo:inherited) and moved off synthetic ids (alias-base-test-4517) onto a real catalog pair (mimo-v2.5 / mimo-v2.5-free), since the base lookup is now catalog-anchored.

Verification

$ npx vitest run --config vitest.mcp.config.ts open-sse/services/autoCombo/__tests__/
 Test Files  7 passed (7)
      Tests  112 passed (112)

$ node --import tsx/esm --import ./open-sse/utils/setupPolyfill.ts \
    --import ./tests/_setup/isolateDataDir.ts --test \
    tests/unit/task-fitness-scores-as-11489.test.ts tests/unit/task-fitness-missing-table-8603.test.ts
# tests 5
# pass 5
# fail 0

$ npx eslint <9 touched files>
✖ 3 problems (3 errors, 0 warnings)
  open-sse/config/providers/shared.ts  784:51  'model' is defined but never used
  open-sse/services/autoCombo/__tests__/autoCombo.test.ts  5:44 'vi' / 7:34 'ScoringWeights' unused

All 3 are pre-existing on the base branch (#11449) and are already frozen in config/quality/eslint-suppressions.json with exactly these counts (shared.ts: 1, autoCombo.test.ts: 2), which is what npm run lint (and CI) applies. 0 new errors.

$ npm run typecheck:core
open-sse/services/compression/engines/omniglyphAdapter.ts  (6 errors)
open-sse/services/compression/omniglyphTelemetry.ts        (3 errors)

Both are pre-existing omniglyph package-export mismatches in files this PR does not touch; no touched file reports an error.