fix(sse): route task-aware defaults by intent, fix fitness pattern shadowing (#8602, #8603)

[MumuTW]

Fixes #8602, fixes #8603.

Two defects in the hand-maintained model-quality tables that the auto-router leans on.

#8602 — task-aware defaults hardcoded stale model ids

coding: "deepseek/deepseek-chat",
analysis: "gemini/gemini-2.5-pro",
vision: "openai/gpt-4o",                    // 2024-05
summarization: "gemini/gemini-2.5-flash",
background: "gemini/gemini-2.5-flash-lite",

Wrong twice over:

• The ids rot. They aged out by a generation or two, and every model release made them staler.
• The shape bypasses the router. applyTaskAwareRouting overwrites body.model (src/sse/handlers/chat.ts:573), so a literal target skips auto-combo's 13-factor scoring (quota, circuit-breaker health, cost, latency, stability), connection cooldown and model lockout. An operator with no OpenAI connection got a vision request rewritten to openai/gpt-4o and a hard failure, where pass-through would have worked.

Refreshing the strings would only reset the rot clock, so the defaults now name intents:

task	before	after
coding	deepseek/deepseek-chat	auto/coding
analysis	gemini/gemini-2.5-pro	auto/reasoning
vision	openai/gpt-4o	auto/vision
summarization	gemini/gemini-2.5-flash	auto/chat:fast
background	gemini/gemini-2.5-flash-lite	auto/chat:cheap

creative and chat stay pass-through. These resolve on demand against the operator's actually-connected backends (suffixComposition.ts → virtualFactory.ts) and degrade gracefully as backends rotate. No provider/model literal remains in the module — the stale header comment is updated too.

Operators who pinned a specific model via PUT /api/settings/task-routing are unaffected; only the shipped defaults change. The feature remains off by default.

#8603 — fitness table matched the wrong row

lookupStaticFitnessTable returned the first String.includes hit in declaration order. FITNESS_TABLE.coding declares "gpt-4o": 0.9 before "gpt-4o-mini": 0.8, so an explicitly cheap model inherited the flagship's task fitness and its own row was unreachable:

gpt-4o-mini    -> matched "gpt-4o"      = 0.9   (own row = 0.8)
deepseek-v3.2  -> matched "deepseek-v3" = 0.85  (own row = 0.86)

Now matches longest-pattern-first, so the most specific row wins regardless of authoring order and the table can be extended without ordering hazards. Extracted as the exported getStaticFitnessTableScore — the surrounding resolution chain queries the DB (user_override / arena_elo / models.dev tier) before reaching layer 4, so testing through getTaskFitness would depend on DB fixture state.

Bounded but real: taskFit is weight 0.08 in DEFAULT_WEIGHTS, 0.37 in the quality-first mode pack, and layer 4 is what the long tail of the provider catalog actually lands on.

Validation (Hard Rule #18 — TDD)

Both written first and confirmed red.

tests/unit/task-router-auto-intent-8602.test.ts (node runner) — 4 cases, red before:

expected an auto intent id, got "deepseek/deepseek-chat"

open-sse/services/autoCombo/__tests__/taskFitness-pattern-order-8603.test.ts (vitest) — 7 cases, red before:

TypeError: getStaticFitnessTableScore is not a function

Green after, plus the full blocking vitest job:

node --import tsx/esm --test tests/unit/task-router-auto-intent-8602.test.ts   # 4/4
npm run test:vitest                                # Test Files 32 passed / Tests 281 passed

The #8602 guard is structural, not a value check — it asserts no default may be a concrete provider/model literal, so the rot cannot come back.

Gates

• npm run typecheck:core — clean
• npm run test:vitest — 32 files / 281 tests, all pass
• npx eslint on changed files — 4 no-explicit-any in taskAwareRouter.ts, all pre-existing (eslint-suppressions.json records count: 4 for this file); taskFitness.ts clean. No new violations.

Not in scope

The detection patterns (taskAwareRouter.ts:46-155) are English-only literal substrings, so non-English prompts always fall through to chat. Noted in #8602, deliberately not addressed here. The static FITNESS_TABLE should eventually be deleted rather than curated once arena_elo/models.dev coverage is measured — noted in #8603.

[MumuTW]

The two red checks are pre-existing base breakage, not this PR

Fast Quality Gates and Unit Tests fast-path (2/4) fail identically on the base commit itself. Verified by running both on a pristine worktree at 4053e2314 with no changes from this PR:

Check	Pristine base 4053e2314	This PR's CI
check:complexity-ratchets	complexity=2169 / cognitiveComplexity=956	complexity=2169 / cognitiveComplexity=956
tests/unit/serial/combo-quota-share-cooldown-wait-timing.test.ts	2 pass / 2 fail — 429 !== 200, 403 !== 429	same two assertions

Byte-identical. Unrelated PR #8599 fails on the same two jobs in the same CI window.

Already tracked, no new issue filed:

• fix(quality): release branch not green — DB rules violation in compressionDetailNormalizers.ts #8540 — release branch not green; lists these exact numbers (956 vs baseline 951, complexity=2169) and classifies them as "Ratchet drift (non-blocking — rebaseline at release) … NOT a contributor concern"
• fix(resilience): combo cooldown-wait gate still quota-share/auto only, contradicting its own "ALL strategies" contract (#7301) #8541 — the failing combo-quota-share-cooldown-wait-timing cases
• fix(quality): absolute baselines make inherited drift red every PR — red stops being a signal #8522 — absolute ratchet baselines make inherited drift red on every PR
• fix(ci): base-red compounds because fast-gates fail-fast hides later gates, and PRs land while red #8542 — fast-gates fail-fast hides later gates

Per #8542, a green board can't be the evidence here — the fail-fast masks downstream gates. Local validation for this PR:

node --import tsx/esm --test tests/unit/task-router-auto-intent-8602.test.ts
# tests 4 / # pass 4 / # fail 0

npm run test:vitest        # Test Files 32 passed / Tests 281 passed
npm run typecheck:core     # clean
npx eslint <changed files>  # 4 no-explicit-any in taskAwareRouter.ts, all pre-existing
                            # (suppressions record count: 4); taskFitness.ts clean

Everything else on this PR is green: Vitest, Unit 3/4, ESLint, Docs Gates, Merge integrity, semgrep.