feat(llmhubs): support tool calling and structured output in the Gemi…

[xinkuleee]

Summary

Adds Google Gemini function/tool calling and JSON-schema structured output to the
LLMHub Gemini adapter, so Gemini models can drive the chat agent (tools) and
structured routing (intent check).

• Tool calling: map internal (OpenAI-shaped) tools / function_call /
  function_result to Gemini functionDeclarations / functionCall /
  functionResponse (role mapping assistant→model / tool→user, and
  call_id↔function-name correlation since Gemini responses carry no call id);
  parse Gemini functionCall back into function_call outputs.
• Schema sanitizer: rewrite tool/response JSON Schema into Gemini's
  OpenAPI-subset — inline $ref/$defs, drop additionalProperties/$schema,
  flatten single-element allOf, and convert Optional anyOf + {"type":"null"}
  to nullable.
• Structured output: map response_format.type=json_schema to Gemini
  responseMimeType + responseSchema (fixes chat intent-check routing on Gemini).
• Capability flags: advertise supports_tools / supports_structured_output
  for the Gemini provider (core io_profile derivation + backend equivalent).
• Tests: unit coverage for request building, input/response mapping, the
  sanitizer (refs/defs/additionalProperties/allOf/anyOf-null/recursion), and
  structured output.

Validation

• make precommit-run
• cd backend && go test ./...
• cd core && uv run pytest
• Frontend build/lint, if working from a frontend source checkout
• Not run; explain below

Frontend is untouched, so its build/lint does not apply.

Checklist

• Linked relevant issues or explained why there is no issue.
• Updated docs and examples where behavior changed.
• Updated CHANGELOG.md under ## [Unreleased] for user-facing changes.
• Regenerated and committed generated files after proto, Wire, or OpenAPI changes.
• Confirmed no secrets, tokens, credentials, or sensitive logs are included.

Notes for reviewers

• Streaming is not implemented for Gemini yet — it falls back to the base
  non-streaming path (generate_stream yields a single chunk). Tool calling works,
  but without incremental token streaming; a follow-up can add
  streamGenerateContent?alt=sse.
• The schema sanitizer targets the keywords Gemini currently rejects; verified
  against the full builtin tool set and the structured intent-check schema.
• No secrets/config included: the local Gemini model YAML (with API key) and the
  local default-model switch are gitignored and intentionally excluded from this PR.

[jasonchen922]

the new allowlist filters schema field names, but it does not validate the value shape required by Gemini's OpenAPI Schema.

In particular, Schema.enum accepts strings only, while the current const conversion emits the original JSON value unchanged. The official google-genai types.Schema validator rejects all of these sanitized Pydantic schemas:

• Literal[3] -> {"type": "integer", "enum": [3]}
• Literal[True] -> {"type": "boolean", "enum": [true]}
• Literal[None] -> an enum containing null
• IntEnum -> an integer-valued enum

The newly added numeric-const test therefore asserts an output that Gemini's Schema contract rejects. The same value-shape issue also applies to type: a valid JSON Schema such as {"type": ["string", "null"]} currently passes through unchanged, but Gemini's Schema.type accepts only one enum value.

Please validate and normalize field values in addition to allowlisting keys. Non-string enums should either be represented using supported constraints, deliberately dropped/rejected, or sent through the JSON-Schema-specific API fields (responseJsonSchema / parametersJsonSchema). Nullable type arrays should be converted to a scalar type plus nullable, where representable. Please add contract tests covering numeric/bool/null literals, IntEnum, and type arrays.