feat(groq): built-in-tools request fork + boundary gating (#69 S4)

[stackbilt-admin]

Implements S4 of the Groq built-in-tools sprint (#69), on top of the merged S3 catalog (#71).

What

Forks the Groq adapter's request shape by model family and gates built-in tools at the boundary.

Model family	Wire shape	Vocabulary
groq/compound, groq/compound-mini	compound_custom.tools.enabled_tools: [...]	normalized identifiers verbatim (compound's vocabulary is the normalized set)
openai/gpt-oss-120b	OpenAI-style tools: [{ type }], merged onto the same array as function tools	translated — web_search → browser_search

Gating reuses the catalog — no drift

The capability gate calls modelSupportsBuiltInTools(model, 'groq', tool.type) (the accessor S3 built) as the single source of truth, rather than a duplicate capability set in the adapter. Only the wire-translation map (web_search→browser_search) is local, since spelling isn't a catalog concern. The enforcement point is still the adapter (satisfies the issue review's "gating belongs in the adapter, not catalog-only" — the accessor takes (model, provider) explicitly, so pinned models that bypass routing are still gated).

• Unsupported (model, tool) → ConfigurationError naming the capable models and, for gpt-oss, its supported subset (web_search, code_interpreter).
• Gate-passed-but-unmapped tool → explicit internal error rather than emitting {type: undefined}.

Also required — the models array

base.validateRequest rejects any pinned model not in the provider's models list before formatRequest runs, so model: 'groq/compound' would have thrown without this. Added both compound models there + matching getModelCapabilities() entries (token costs estimated; built-in surcharges not token-tracked, per issue scope).

Deliberately not touched

• TOOL_CAPABLE_MODELS — compound's function-tool support is unverified from the S0 spike, so the conservative reject stands (built-in tools are a separate, gated path).
• GROQ_RESPONSE_SCHEMA — response parsing is S5. Confirmed validateSchema ignores undeclared fields, so the unchanged schema already passes a real executed_tools response.

Tests

typecheck clean; 398 pass (+7 new): compound fork shape, five-tool verbatim passthrough, gpt-oss translation, function+built-in merge on one array, function-only-model rejection (names capable models), unsupported-tool-on-gpt-oss rejection (names subset), pinned-groq/compound accepted through validateRequest.

Next: S5 — parse executed_tools[].search_results.results → metadata.builtInToolResults, extend GROQ_RESPONSE_SCHEMA.

🤖 Generated with Claude Code