fix(intelligence): synthesize base-model arena rows from effort/harness variants; drop MODEL_ALIAS_MAP (#11504)

[MumuTW]

Fixes #11504

Stacked on #11492 — the diff includes #11492's commits until it merges; only the src/lib/arenaEloSync.ts + tests/unit/arena-elo-sync.test.ts changes are this PR's.

Problem

arenaEloSync stores leaderboard rows under the raw Arena entry name, while task-fitness layer 2 looks rows up by the request's model id. Arena scores harness × effort combinations as separate entries, so the two never meet for a bare id: on a synced DB, 9 of the arena ids carry a (codex-harness)-style annotation and 24 effort variants have no bare-id twin at all. A request for gpt-5.6-sol, claude-opus-5, gemini-3.6-flash or deepseek-v4-flash misses layer 2 with certainty and falls through to the static table.

Same file, same class of defect: MODEL_ALIAS_MAP (8 hand-maintained entries) copied a score under other names, and some of those collapse generations — gpt-5.5 → gpt-5, claude-opus-4-6-thinking → claude-opus-4, kimi-k2-thinking → moonshot/kimi-k2.

Mechanism

normalizeModelName() only lowercased and stripped a vendor prefix, so gpt-5.6-sol-xhigh (codex-harness) was stored verbatim — even a request for the exact variant id missed. And nothing ever produced the base id, because the leaderboard never lists it.

Fix

1. normalizeModelName() drops a trailing (…) harness annotation (after lowercasing, before the vendor-prefix strip). The exported contract is otherwise unchanged.
2. The variant row is kept as-is — it is a real routable id for several providers.
3. A new pass after the per-model loop (withSynthesizedBaseRows) adds one synthesized base row per (base, category) for every entry that resolveScoresAs() (fix(autoCombo): inherit task fitness from base model for effort/alias variants (#11489) #11492) resolves to a routable catalog id: score = the best task fit among that base's variants, eloRaw and confidence from the winning variant (no invented confidence label — the column's vocabulary stays high/medium/low). Rationale: the per-effort entries measure the same weights at different budgets, so the model's ceiling is the max, and cost/latency are already separate factors in the 12-factor score.
4. A base the leaderboard measured directly is never lowered: the synthesized candidate replaces it only when strictly higher.
5. MODEL_ALIAS_MAP and its expansion loop are deleted. Alias coverage is now scoresAs + RegistryModel.aliases, which the registry owns. No replacement table.
6. Resolution never guesses: resolveScoresAs returning via: null (e.g. grok-4.6-fast-high → grok-4.6-fast is not a catalog id) synthesizes nothing.

No schema change; bulkUpsertModelIntelligence is INSERT OR REPLACE on (model, source, category).

Measured effect

Fixture built from the real distinct arena ids on a synced DB (SELECT DISTINCT model, elo_raw FROM model_intelligence WHERE source='arena_elo' AND category='coding'), fed through transformToModelIntelligence in a DATA_DIR-isolated script:

distinct arena ids in	50 (9 of them harness-annotated)
coding rows out	64
synthesized base rows	14

New bases now reachable by a bare request: claude-opus-5, claude-sonnet-5, deepseek-v4-flash, gemini-3.5-flash, gemini-3.6-flash, gemini-3.7-flash, glm-5.2, glm-5.3, gpt-5.4, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, grok-4.6, kimi-k3.

Tests

TDD (Hard Rule #18): the new describe("transformToModelIntelligence() — base-model synthesis") block was written first and failed 5/5 on the pre-fix code, then passed.

• normalizeModelName("gpt-5.6-sol-xhigh (codex-harness)") → gpt-5.6-sol-xhigh; same with a vendor prefix and mixed case.
• gpt-5.6-sol-xhigh (codex-harness) → rows for the variant and gpt-5.6-sol; no gpt-5.6 row (that is a vendor alias, resolved at lookup time).
• claude-opus-5-high (1663) + claude-opus-5-max (1691) → the claude-opus-5 row carries the -max score and eloRaw 1691.
• Leaderboard listing both claude-opus-5 (1700) and claude-opus-5-max (1691) → the base keeps 1700 (one row, explicit measurement not lowered).
• gpt-5.5 never emits a gpt-5 row (regression guard for the deleted alias map).
• grok-4.6-fast-high synthesizes nothing — asserted against resolveScoresAs(...).via === null rather than a frozen catalog fact, so the test stays true as the catalog moves.
• Idempotence: two calls on the same input are equal, and (model, category) keys are unique.

Two pre-existing tests were deliberately inverted, not deleted. Both asserted MODEL_ALIAS_MAP expansion — i.e. exactly the generation collapse this PR removes (a claude-opus-4-6-thinking score copied onto claude-opus-4). They now stand as anti-collapse regression guards on the same fixture:

was	is now
"expands model aliases for known models" — asserts claude-opus-4 / anthropic/claude-opus-4 ARE emitted	"does not copy a variant's score onto a different generation (MODEL_ALIAS_MAP removed, #11504)" — asserts claude-opus-4-6-thinking is emitted and neither alias is
"model aliases are stored in DB alongside canonical names" — asserts a claude-opus-4 DB row exists after syncArenaElo()	"stores no cross-generation alias rows in the DB (MODEL_ALIAS_MAP removed, #11504)" — same run, asserts that row is absent

Every other pre-existing test is untouched.

Verification commands

DISABLE_SQLITE_AUTO_BACKUP=true node --import tsx/esm --import ./open-sse/utils/setupPolyfill.ts --import ./tests/_setup/isolateDataDir.ts --test tests/unit/arena-elo-sync.test.ts
# tests 56 | pass 56 | fail 0

npx eslint src/lib/arenaEloSync.ts tests/unit/arena-elo-sync.test.ts --suppressions-location config/quality/eslint-suppressions.json
# 0 errors

npm run typecheck:core
# clean

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