update-models-nvidia

Update src/modules/llms/server/openai/models/nvidianim.models.ts with the latest model definitions for NVIDIA's hosted endpoint (integrate.api.nvidia.com / build.nvidia.com).

Reference src/modules/llms/server/llm.server.types.ts and src/modules/llms/server/models.mappings.ts for context only. Focus on the model file, do not descend into other code.

Primary source - run the harvest tool (requires NVIDIANIM_API_KEY in .env.api-keys for the probe phase; never commit or echo the key):

npx tsx tools/develop/nvidianim-catalog-sync/harvest.ts

This merges 4 sources (see the tool's README): live /v1/models ids, build.nvidia.com markdown catalog (labels, capabilities, createdDate -> pubDate), NGC deprecation dates, and authenticated liveness + context-window probes. It writes tools/develop/nvidianim-catalog-sync/harvest-latest.json. The probe phase is paced for the 40 RPM account limit and takes 30-45 minutes; use --skip-probes for a metadata-only refresh.

harvest-latest.json is a committed snapshot (id-sorted, diff-stable): after the run, git diff tools/develop/nvidianim-catalog-sync/harvest-latest.json IS the change review - alive flips, context changes, and new deprecation dates in that diff are exactly the edits to port into the curated table. Commit the refreshed snapshot together with the table changes. Ignore harvestedAt and lastMonthInvocations churn.

Do NOT use web search. The harvest output is the ground truth.

Applying the results to the curated table:

• ADD newly-alive chat models worth surfacing (skip embeddings/rerankers/parsers/guards unless hidden)
• REMOVE models that are dead (dead-entitlement, retired, no-chat-route) or carry a deprecationDate in the past or within days
• REGENERATE _retiredNvidiaNIMIds (the deny list): every harvested id that is neither curated nor a live chat model belongs there, grouped by classification. The invariant is total coverage: curated + denied should equal the full /v1/models list, so that any id outside both sets is a genuine 0-day arrival (those surface automatically as hidden entries)
• CAUTION - dead-for-our-key is not dead-for-everyone: NVIDIA scopes function visibility per account, so dead-entitlement and probe-error classifications reflect OUR key only. Before denying such a model, cross-check production analytics (PostHog, host integrate.api.nvidia.com, successful aix_chat_generate_completed events in the last ~14 days) for other accounts using it successfully. Precedent: qwen/qwen3.5-397b-a17b probed dead for our key on 2026-07-25 yet had 17 recent successes from 4 users - it stays OFF the deny list (hidden 0-day entry). Models with zero successes across all accounts are safe to deny
• contextWindow MUST come from the measured probe value (ctxMeasured), never from build.nvidia.com's advertised value - they disagree on ~25% of models, up to 8x, and gemma-4-31b silently truncates
• pubDate is the upstream model release date: prefer the same model's pubDate from another vendor's *.models.ts (add a // = <file> '<id>' cross-reference comment), fall back to the harvest pubDate (catalog createdDate)
• Borrowed benchmark: { cbaElo } values use the - 2 yield idiom so native vendors win auto-picks
• Keep chatPrice: _freePrice on all models (the endpoint has no paid tier)
• Reasoning params: gpt-oss models use _PS_OaiEffort; other thinking models use _PS_Thinking (wired to chat_template_kwargs in the adapter)
• Preserve comments and table order (flagships first, hidden tail last); minimize whitespace churn

Verify: tsc --noEmit --pretty && npm run lint, then NVIDIANIM_API_KEY=... npm test (the nvidianim tests do a live listing; a [DEV] stale warning fails the test - that is the drift alarm).