serve base_model in api.json and catalog.json

[arimxyer]

What

Provider model TOMLs already declare base_model — the edge linking an offering to its canonical models/ registry entry — but generate.ts strips the field during inheritance merging, so neither api.json nor catalog.json carries it. Consumers who want to group provider offerings by underlying model (e.g. "which providers serve anthropic/claude-opus-4-6, and at what price?") have to re-derive that linkage themselves from heuristics, even though the repository already knows the answer.

The site itself pays for this today: loadProviderBaseModelRefs in packages/web/src/render.tsx re-parses every provider TOML a second time solely to rebuild the refs the generate pipeline just dropped.

This PR retains the resolved ref on the emitted model instead:

• generate.ts keeps base_model on the provider model after the inheritance merge validates (the ref is guaranteed to resolve at this point — mergeBaseModel throws otherwise)
• ModelShape allows the field as an output-only optional; authored TOMLs still declare it via the BaseModel wrapper schema, and models.json (registry metadata) is unaffected
• render.tsx derives BaseModelRefs from the generated catalog and drops the duplicate TOML scan
• the SDK Model type documents the field (generated effect schema updated via bun run generate)

Net diff is ±25 lines.

Why

base_model is the only provider-agnostic identity signal in the dataset, and it's currently invisible to every API consumer. Serving it makes catalog.json self-contained for cross-provider grouping: group_by(model.base_model ?? id) replaces per-consumer name/id heuristics. It also removes the site's own double-parse workaround.

Verification

• bun validate passes; 2,148 refs surface in _api.json / _catalog.json
• the catalog-derived BaseModelRefs map is identical to the previous raw-TOML scan: 2,148/2,148 entries, 0 mismatches (so site grouping behavior is provably unchanged)
• models.json carries no new field (0 occurrences)
• SDK drift suite passes: bun run generate && tsc --noEmit && bun test (23 pass)

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] packages/core/test/generate.test.ts:38 - Check: Catalog tests must match the new base_model emit contract. Why: Two tests still require provider JSON to omit base_model: base_model can factor metadata without changing provider JSON (equality + not.toHaveProperty("base_model")) and repository provider JSON strips authored metadata pointers (fails if encoded output contains base_model). This PR keeps the ref on emitted models, so those assertions are now wrong and will fail under bun test. Action: Rewrite them to assert resolved base_model is retained, factored vs inline models match after omitting that field, and only base_model_omit (plus other non-output pointers) stays stripped.
• [medium] [violation] AGENTS.md:103 - Check: AGENTS.md is authoritative for base_model output behavior. Why: It still says the final provider JSON contains no base_model or base_model_omit fields, which this change deliberately reverses for base_model. Leaving that line would document the opposite of the served API. Action: Update that sentence to state resolved base_model is emitted on provider models and base_model_omit remains parse-only / never served.

[arimxyer]

Updated the branch to also bring the core test suite along with the contract change:

• the factored-fixture test now expects the retained ref ({ ...direct, base_model: "lab/model" }) instead of asserting its absence
• the repository-wide leakage test now asserts base_model_omit never leaks and that every served base_model resolves to a models/ registry entry (a stronger guarantee than before)

Note: bun test has one unrelated failure on current dev (repository open-weight model metadata includes weights links — sarvam/sarvam-105b has open_weights = true but no weights links). It pre-exists this branch; happy to fix it separately.

[arimxyer]

Both review action items are addressed: the test contract was updated in the previous push (factored fixture expects the retained ref; the repo-wide test now guards base_model_omit only and additionally asserts every served ref resolves), and AGENTS.md now documents that resolved base_model is served while base_model_omit stays authoring-only.

[github-actions]

No actionable findings.