feat(scripts): sync the full models.dev fact set (pricing, description, knowledge, input limits) #2329

Problem

The models.dev pipeline (scripts/sync-model-metadata.mjs) downloads https://models.dev/api.json (~3.5 MB, 6149 models) on every sync, but only maps a fraction of the data. Several high-value facts are dropped:

• Pricing: packages/runtime/src/telemetry/builtin-pricing.ts is a hand-maintained 17-entry snapshot ("as of 2026-05-20"); models.dev carries cost on 93% of models (5739/6149). Spot check: local claude-sonnet-4-5 (3 / 15 / 0.3 / 3.75) matches models.dev exactly — the snapshot was transcribed by hand. No provider exposes a machine-readable pricing API, so models.dev is the correct source, and it is already a dependency.
• Description: present on 100% of models, never synced — the model picker shows only a display name, so users cannot tell same-family models apart.
• Knowledge cutoff: present on 53% (3271) of models, never synced — users see no "knowledge as of" anywhere.
• Input limit: limit.input exists on 19% (1195) of models (e.g. gpt-5.2: context 400K but input 272K). Only context and output are synced today; the context budget (context-budget-policy.ts:410-424) budgets against contextWindow, which can over-admit input the model cannot actually take.
• PDF attachment capability: modalities.input values include pdf (e.g. claude-sonnet-4-5), but the sync filter keeps only text/image/audio — the pdf fact is dropped (relevant to attachment preflight).
• Structured output: present on 62% (3829) of models, never synced — distinct from tool_call (function calling); a model can support tools without structured JSON output.
• Freshness: last_updated present on 100% of models, never synced — generated facts carry no age marker, so stale entries rot silently (the exact failure mode of the 17-entry price snapshot).
• Status nuance: models.dev status has deprecated (170), beta (53), alpha (1); the sync maps only deprecated → deprecated, everything else → active — beta/alpha models are presented as stable.

Proposed shape (extend the existing seam, no new pipeline)

1. Extend sync-model-metadata.mjs to map the full fact set into generated modules:
   • cost → model-pricing.generated.ts (base rates only; see non-goals)
   • description, knowledge, limit.input, structured_output, last_updated → extend ModelMetadata / model-metadata.generated.ts
   • modalities.input keeps pdf (alongside text/image/audio)
   • status: map beta/alpha distinctly instead of collapsing into active
2. builtin-pricing.ts becomes generated + a local supplement layer for entries models.dev cannot carry (e.g. zai-coding-plan:*, MiniMax-cn:* — special tiers models.dev does not price). Keep getBuiltinPricing / buildPricingLookup signatures stable (runtime, runtime-host, desktop all consume them).
3. Extend the sync-contract test pattern (models-dev-sync-contract.test.ts truthfulness checks) to the new fields.

Non-goals

• Tiered pricing (OpenAI tiers / context_over_200k): the flat PricingConfig takes the base rate only.
• Special tiers (Coding Plan, Education, Volume Pricing) stay overrides, matching the current comment's intent.
• Low-value fields (family, open_weights, release_date, temperature, interleaved, experimental, provider-level api/env) are intentionally not synced.

Impact

• Removes the silently-stale hand-maintained price snapshot.
• Model picker can show description + knowledge cutoff; budget decisions respect real input limits; attachment preflight gets the pdf fact; beta/alpha models are honestly marked.
• Complements feat: user-overridable model facts (context window etc.) — models.json-style override layer #2330 (the user override layer): generated facts = read-only built-in side, models.json = user side.

Validation

• Contract test asserting generated pricing/description/knowledge/limit entries match api.json.
• A diff test that the current 17 hand entries still resolve through the new lookup.
• Context budget test: a model with limit.input budgets against the input limit, not contextWindow.