feat(providers): curated OpenRouter embeddings catalog + specialty merge in live discovery (#6976)

[diegosouzapw]

Closes #6976

Root cause

OpenRouter serves embeddings via a dedicated OpenAI-compatible /api/v1/embeddings endpoint (omitted from /v1/models). open-sse/config/embeddingRegistry.ts already had an openrouter entry, but it was stale (3 legacy ids) and effectively dead code for discovery: providerModelsConfig.ts gives openrouter a live discovery config, so the live-discovery success path in buildApiDiscoveryResponse (src/app/api/providers/[id]/models/route.ts) returned the live /v1/models chat catalog verbatim — the specialty (embeddings/rerank) static catalog was only ever merged in on the no-config local_catalog fallback, which OpenRouter never hits.

Fix

1. Refreshed the curated openrouter embeddingRegistry lineup — ids verified against the API reference (https://openrouter.ai/docs/api/reference/embeddings) and the models collections page, not display names: openai/text-embedding-3-small/-large, qwen/qwen3-embedding-8b/-4b, baai/bge-m3, mistralai/mistral-embed-2312, google/gemini-embedding-001. Dropped the legacy openai/text-embedding-ada-002 entry rather than guess at its current availability.
2. Added mergeSpecialtyCatalogIntoLiveModels() (discovery/helpers.ts) that folds embeddings/rerank entries from getStaticModelsForProvider() into a successful live-discovery response, additively and deduped by id. Scoped via an explicit allowlist (LIVE_DISCOVERY_SPECIALTY_MERGE_PROVIDERS, currently just openrouter) rather than applied to every provider with an embeddingRegistry/rerankRegistry entry — some providers (Gemini) already return embedding models directly inside their live /v1/models response, so a blanket merge broke an existing pagination test (gemini-embedding-2/gemini-embedding-001 duplicating what Gemini's own live response already returns). Updated one pre-existing OpenRouter test (provider-models-route.test.ts) whose exact-equality assertion encoded the old (buggy) behavior.

TDD evidence

New tests/unit/openrouter-embeddings-catalog-6976.test.ts:

• RED (Step 2/merge reverted): live discovery merges curated embeddings into the response even when /v1/models returns none failed — AssertionError: curated embedding baai/bge-m3 should be merged into live discovery; got: anthropic/claude-sonnet-5.
• GREEN (fix applied): all 4 new tests pass (registry lineup + dimensions, getStaticModelsForProvider fold, live-discovery merge, live-vs-curated dedup-by-id).

Gates run

• npm run typecheck:core — clean
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — 0 errors (pre-existing no-explicit-any warnings only, all suppressed)
• node scripts/check/check-file-size.mjs — no new violations (3 pre-existing base-red files untouched)
• node scripts/check/check-complexity.mjs / check-cognitive-complexity.mjs — both OK at baseline (2056 / 890, unchanged)
• node scripts/check/check-changelog-integrity.mjs — OK
• Full tests/unit/provider-models-route.test.ts (59 tests), provider-models-discovery-split.test.ts, provider-models-custom-merge-6247.test.ts, provider-models-route-codex.test.ts, provider-models-v1-route.test.ts, provider-scoped-models-route.test.ts, openrouter-registry.test.ts, rerank-openrouter-6574.test.ts — all green (regression sweep of the area)

Notes

• Left nvidia/llama-nemotron-embed-vl-1b-v2 and perplexity/pplx-embed-v1-0.6b (both surfaced in the API reference/collections search) out of the curated list — less confident these fit the plain OpenAI-shaped embeddings request/response contract without further verification (multimodal / different input shape); flagging for a follow-up refresh rather than guessing.
• google/gemini-embedding-2 (128–3072 flexible dims per the collections page) also left out — no single fixed dimension to record for the conflict guard, unlike gemini-embedding-001 (768, matching the existing gemini registry entry's convention in this file).

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[diegosouzapw]

PR #6994 — aprovado. Probe TDD confirmou fail-without-fix no teste-chave do merge (revertendo route.ts/helpers.ts ele cai pra RED com a mensagem exata do bug original). Fui atrás dos IDs curados no OpenRouter (mistral-embed-2312, bge-m3, qwen3-embedding-4b/8b) — todos existem e batem com as dimensões documentadas. Único ponto de atenção (não bloqueante, não é regressão sua): google/gemini-embedding-001 fica registrado com dimensions: 768 mas o default real da API do Google pra esse modelo é 3072 — só que isso já é a convenção usada no entry nativo gemini no mesmo arquivo (linha 221), então você só espelhou um comportamento pré-existente. Vale um follow-up separado se quiser corrigir os dois de uma vez. Escopo do allowlist (só openrouter) bem justificado — evita duplicar embeddings no Gemini, que já retorna isso nativamente.

[diegosouzapw]

Babysit summary

CI is green (13/13 SUCCESS, 2 NEUTRAL = Mergify skips). Not merged — handing off for human review & merge.

Fixes

#	Commit	Gate	Rationale
1	6d0f8e23b	Fast Quality Gates, Unit Tests (1–4/4), Vitest	Merged release/v3.8.49 in. The branch was 57 commits behind; the stale base carried an orphan test (open-sse/translator/request/__tests__/openai-to-gemini.test.ts) already relocated on the release tip, which tripped check:test-discovery. The unit/Vitest reds were runner-infra SQLITE_FULL flakes, gone on re-run. Clean merge, no conflicts.
2	3ac194f76	No new ESLint warnings	The new openrouter-embeddings-catalog-6976.test.ts used any 4×; no-explicit-any is an error under tests/ (#6218). Replaced with an explicit ModelsResponseBody type.
3	6b0379249	No new ESLint warnings	The #6976 assertion added a 56th explicit any to provider-models-route.test.ts, one over the 55 frozen in config/quality/eslint-suppressions.json. Typed the callback param rather than raising the frozen count — the debt ratchet only decreases.

Guardrails honored

• No assertion weakened or removed — assert counts unchanged (13 → 13 and 300 → 300); both suites still fully green (4/4 and 59/59).
• No CI definition, ratchet baseline, or suppressions file edited. eslint-suppressions.json deliberately left at 55 — the violation was fixed at the cause.
• Complexity/cognitive ratchets verified locally: complexity=2056 (baseline 2056), cognitiveComplexity=890 (baseline 890) — no regression.
• Full local lint gate: 0 errors, 0 warnings.

Note on the base: release/v3.8.49's own CI is red on Quality Gates (Extended), Integration Tests (2/2), Package Artifact, Electron Package Smoke and Quality Ratchet. Those are pre-existing base-drift, unrelated to this PR, and are not in this PR's check set.

Ready for human review & merge.