feat(google): support OpenAI-compatible reasoning_effort for Gemini thinking models

[Tushar49]

Adds OpenAI-compatible reasoning_effort parameter for the Google provider, mapping to Gemini's thinkingConfig so callers can opt into Gemini's thinking budget without leaving the OpenAI request shape.

What this PR does

• Schema: shared/types.ts + server/src/routes/proxy.ts accept reasoning_effort: "minimal" | "low" | "medium" | "high" on ChatCompletionRequest. Omitting it preserves existing behavior exactly (no thinkingConfig is sent).
• Mapping (server/src/providers/google.ts):
  • Gemini 3.x → thinkingConfig.thinkingLevel (MINIMAL / LOW / MEDIUM / HIGH). gemini-3.1-pro does not accept MINIMAL → falls back to LOW.
  • Other Gemini models (2.x, etc.) → thinkingConfig.thinkingBudget integer. Canonical mapping: minimal → 512 (Pro: 128), low → 1024, medium → 8192, high → 24576. Clamped per model: gemini-2.5-pro min 128 / max 24576, Flash 0–24576, Flash-Lite 512–24576.
  • Always sets includeThoughts: true when reasoning_effort is provided.
• Response parsing: Gemini returns candidates[0].content.parts[] where parts may carry thought: true. Non-thought parts go to message.content (unchanged); thought parts are aggregated into message.reasoning_content (OpenAI extension used by opencode and other clients). usageMetadata.thoughtsTokenCount → usage.completion_tokens_details.reasoning_tokens.
• Streaming: thought-part deltas emit as delta.reasoning_content, normal text deltas as delta.content.

Why

OpenAI clients (opencode, Continue, Cursor, etc.) already use reasoning_effort to drive reasoning models. Today they have no way to dial Gemini's thinking budget through freellmapi without bypassing the proxy. Google's own OpenAI-compatible endpoint maps the same field to the same thinkingConfig — this PR brings freellmapi to parity.

Verification

From server/:

• bun run build (tsc): clean
• bun run test src/__tests__/providers/google-reasoning.test.ts: 8 / 8 pass (1 file)
• bun run test src/__tests__/providers/google.test.ts src/__tests__/providers/google-schema.test.ts: 19 / 19 pass (no regressions to existing Google tests / PR fix(google): strip every JSON Schema key not in Google's Schema proto (examples/const/readOnly/writeOnly/uniqueItems/not/allOf/oneOf/...) #105 tests)

New tests cover: effort → budget mapping per model class, Gemini 3.x thinkingLevel path, gemini-3.1-pro MINIMAL → LOW fallback, includeThoughts on/off semantics, thought-part parsing into reasoning_content, reasoning_tokens in usage, and streaming delta.reasoning_content.

Style / scope

• Matches the four-commit conventional-commit shape of fix(google): strip every JSON Schema key not in Google's Schema proto (examples/const/readOnly/writeOnly/uniqueItems/not/allOf/oneOf/...) #105 (feat(api): → feat(google): → feat(proxy): → test(google):).
• No new dependencies. No request-schema removals or renames.
• When reasoning_effort is omitted, generationConfig.thinkingConfig is not added to the outgoing request, so existing callers see byte-for-byte the same Gemini payload.

Happy to iterate on naming, default budgets, or split into smaller PRs if preferred.