fix(vercel): fill non-language model costs from gateway pricing

[rekram1-node]

Summary

• Vercel sync previously only mapped chat-style input+output token pricing, so embed/image/video/audio models often had no [cost] at all.
• Parse token-like pricing when available (including input-only for embeddings/speech/rerank, with output = 0).
• For non-token shapes (per-image, per-second video) or missing pricing, write explicit zero costs and seed a leading header comment that survives sync refreshes until better non-token cost tracking exists.
• Resynced 96 Vercel models so every gateway entry now has a cost block.

Test plan

• bun test packages/core/test/sync.test.ts (including new embed/zeroed pricing cases)
• bun run vercel:generate updates 96 models, then re-run is fully unchanged
• bun validate succeeds
• Confirm a sample embed model has parsed input cost and a sample image/video model has zeros + header

[github-actions]

Action items

• [medium] [possible mistake] providers/vercel/models/zai/glm-4.6v-flash.toml, providers/vercel/models/perplexity/sonar.toml, perplexity/sonar-pro.toml, perplexity/sonar-reasoning-pro.toml (and cohere/rerank-v3.5.toml, rerank-v4-fast.toml, rerank-v4-pro.toml) - Check: Zeroed placeholder cost + "Vercel pricing for this model is not token-based" header must not be applied to token-based models. Why: These files declare token context windows (128_000–200_000) and, for glm-4.6v-flash/sonar-reasoning-pro, reasoning = true + tool_call — i.e. they are unambiguously token-priced LLMs. The seeded header ("not token-based … per-image, per-second video") is directly contradicted by the model's own [limit] context in the same file, and input = 0, output = 0 surfaces in api.json as "free," which is unsupported for premium search/reasoning LLMs and conflicts with their known per-token pricing. The contrast with voyage/rerank-2.5*.toml (same model class, but receiving real token input prices) shows the header conflates "Vercel API returned no pricing" with "non-token model." Action: Verify against the Vercel gateway API which of these are genuinely non-token-priced vs. merely missing API pricing. For token-based LLMs/rerankers, do not write the "not token-based" header and do not emit a false zero cost (preserve absent cost, or source the real price); reserve the zeroed placeholder for true per-image/per-second/per-request models.
• [medium] [possible mistake] packages/core/src/sync/index.ts (header selection, diff lines 135-141) and packages/core/src/sync/providers/vercel.ts (diff lines 169-170) - Check: The seeded "Cost currently zeroed" header must be cleared when the placeholder condition no longer holds. Why: The new logic header = existingHeader || (translated.header ?? "") makes existing headers always win, and vercel.translateModel only sets header when costPlaceholder is true (else undefined). If the Vercel API later exposes real token pricing for a currently-zeroed model, costPlaceholder becomes false, translated.header is undefined, but existingHeader still returns the stale "Cost currently zeroed … not token-based" block — so the rewrite emits that header above a non-zero [cost], producing a self-contradictory file. There is no test covering header preservation/clearing across a placeholder→real transition. Action: Add a clearing path (e.g., have the provider return an explicit empty/sentinel header when not a placeholder and have the runner honor it, or make the header conditional on the current placeholder state rather than unconditionally preserving it) and add a test for the refresh-when-pricing-appears case.
• [low] [possible mistake] PR body / providers/vercel/models/openai/text-embedding-3-large.toml, tts-1.toml, whisper-1.toml, voyage/voyage-3-large.toml, etc. - Check: Material factual pricing changes should cite a first-party source. Why: This PR sets concrete per-million-unit prices for ~30 embedding/TTS/whisper/rerank models (e.g. tts-1 input = 15, whisper-1 input = 0.0001, text-embedding-3-large input = 0.13) plus the zero-cost claims above, but the PR body cites no Vercel gateway endpoint or pricing page. TTS/Whisper pricing is per-character or per-minute on the upstream provider, so the unit mapped into cost.input (documented as "per million tokens") cannot be reviewed from the diff alone. Action: Add a citation to the Vercel AI Gateway models/pricing endpoint (or the relevant provider pricing page) stating what it supports, and clarify the billing unit for the speech/transcription entries so the mapped cost.input values can be verified.