fix(google): map reasoning_effort to thinkingBudget on Gemini 2.5

[aayushbaluni]

Fixes #1639.

reasoning_effort was translated to thinkingLevel for every Gemini model, but thinkingLevel is Gemini 3+ only. On the Gemini 2.5 family the parameter is silently ignored server-side, so reasoning_effort had no effect.

Per Vertex's thinking docs:

Control	Applies to
thinkingLevel (MINIMAL/LOW/MEDIUM/HIGH)	Gemini 3 and newer
thinkingBudget (integer)	Gemini 2.5 family

This is invisible on most Gemini 2.5 variants because they default to thinking-on, so the model reasons anyway and the dropped parameter never shows up. gemini-2.5-flash-lite is the exception: it defaults to a thinking budget of 0, so the request returns HTTP 200 with zero reasoning tokens and no error — exactly what @davidwhite-spec reported.

How it got here

thinkingLevel was added in 318da22, titled "map openai reasoning_effort to gemini thinking level", following a revert of "Thinking level support for Gemini 3 Pro". The mapping was correct for Gemini 3 but applied unconditionally:

if (params.reasoning_effort && params.reasoning_effort !== 'none') {
  generationConfig['thinkingConfig'] = {
    thinkingLevel: params.reasoning_effort,
  };
}

Change

Branch on the model family:

• Gemini 3+ → thinkingLevel (unchanged behavior)
• Gemini 2.5 → thinkingBudget, scaled by effort (minimal 512 → high 24576)
• Unknown/absent model → thinkingLevel, preserving today's default

The helper is exported from google-vertex-ai/transformGenerationConfig.ts and reused by providers/google/chatComplete.ts, which had a byte-identical copy of the bug. The model id is matched with /gemini-2\.5/ so it works for both the bare (gemini-2.5-flash) and provider-prefixed (google.gemini-2.5-flash) forms that reach this transform.

The explicit thinking: {type, budget_tokens} parameter is untouched, and there's a test pinning that.

Tests

New src/providers/google-vertex-ai/transformGenerationConfig.test.ts — 14 tests covering:

• each effort level maps to a positive thinkingBudget on Gemini 2.5, and never emits thinkingLevel
• higher effort produces a larger budget
• all 2.5 variants including the google.-prefixed form
• Gemini 3 still gets thinkingLevel
• unknown model falls back to thinkingLevel
• reasoning_effort: 'none' still emits nothing
• the explicit thinking param is unaffected

Verification

• jest src/providers/google-vertex-ai src/providers/google: 32 passed (18 on unmodified main, so +14 and no regressions).
• Mutation-checked: forcing supportsThinkingLevel back to true (i.e. the pre-fix behavior) fails 7 of the 14 new tests, so they measure the fix rather than decorating it.
• Full jest src/: 61 passed vs 47 on main; the pre-existing suite failures and the two tsc --noEmit errors (bytez/api.ts, open-ai-base/index.ts) are identical before and after this change.
• prettier --check clean on all three files.

Note on scope

The issue also reports that /v1/responses behaves differently from /v1/chat/completions. That routing isn't in the OSS repo — Vertex doesn't register createModelResponse in google-vertex-ai/index.ts — so I can't address the endpoint-axis asymmetry here. This PR fixes the model-axis half, which is the part that lives in this codebase and is wrong on the chat-completions path too. I don't have Vertex credentials, so this is verified by unit test and against Google's documented parameter matrix rather than by a live call; a maintainer with a gemini-2.5-flash-lite key can confirm reasoning tokens become non-zero.