fix(gemini): reasoning_effort="none" disables thinking instead of hiding it

[JamMaster1999]

Description

reasoning_effort="none" (and None) currently maps to ThinkingConfig(include_thoughts=False), which hides thought summaries instead of turning off thinking.

"none" already means off everywhere else in this codebase:

• anthropic sends thinking={"type": "disabled"},
• cohere and deepseek send their disable flags,
• ollama sends think=False,
• messages_compat.py maps Anthropic's disabled back to reasoning_effort="none".

Gemini is the one provider where "none" means "hide but keep billing".

The fix, per path:

• thinking_budget models: send ThinkingConfig(include_thoughts=False, thinking_budget=0), so thinking is actually off.
• Models that reject thinking_budget=0 (e.g. gemini-2.5-pro): Google's own 400 surfaces as the typed error, reacting to the provider's error rather than predicting it with a model list.
  • This is the chosen approach for this repo (OpenAI provider passes reasoning_effort through unconditionally even when the model rejects it combined with tools mozilla-ai/any-llm#1189).
  • The live error is self-explanatory: [gemini] 400 INVALID_ARGUMENT ... "Budget 0 is invalid. This model only works in thinking mode."
• thinking_level models (3.5+): thinking_level has no off value since the lowest tier is MINIMAL.
  • So "none" maps to ThinkingLevel.MINIMAL with a logger.warning, the same clamp-to-nearest convention already used in fix(mistral): clamp reasoning_effort, drop prompt_mode, replay thinking traces mozilla-ai/any-llm#1241 and fix(anthropic): map xhigh to Anthropic xhigh and expose max effort mozilla-ai/any-llm#1108.

Why clamp with warning, instead of error?

On 3.5/3.6 the clamp billed zero thought tokens (table below), so "none" callers get what they asked for. Erroring up front instead would mean guessing from the version string which models those are, the kind of guess mozilla-ai#1189 declined. A model that doesn't support MINIMAL rejects it, and the caller gets the typed 400 (3.7, note below).

The warning is because MINIMAL is the least thinking the model allows, not a promise of zero, and mozilla-ai#1107 treats surprise thinking costs as a bug.

None stays merged with "none" as it has been since mozilla-ai#708; "auto" remains the unset default, so callers who don't set the parameter are untouched.

Verified live across every path with reasoning_effort="none" (gap = thought tokens billed but hidden; before the fix, gemini-2.5-flash billed 424 hidden thought tokens):

gemini models	path	result
2.5-flash, 2.5-flash-lite, 3-flash-preview, 3.1-flash-lite	budget → 0	OK, gap 0
2.5-pro, 3.1-pro-preview	budget → 0, rejected	typed InvalidRequestError: "Budget 0 is invalid. This model only works in thinking mode."
3.5-flash, 3.5-flash-lite, 3.6-flash	level → MINIMAL	OK, gap 0, warning emitted
3.7-flash	level → MINIMAL, rejected	typed InvalidRequestError: "Thinking level MINIMAL is not supported", details below

Every failure surfaces as a typed any-llm exception; no raw google.genai errors leak.

Why gemini-3.7-flash rejects the clamp

gemini-3.7-flash cannot disable thinking: MINIMAL is gone, the new floor LOW thinks for real (220 thought tokens on the test prompt), and thinking_budget=0 is silently ignored but still billed. So the 400 is correct, same as the thinking-only pros, and clamping to LOW would just bring back the bug this PR fixes. The removal also breaks the existing "minimal" mapping (mozilla-ai#1281) on 3.7.

For reference, thinking_budget=0 today is honored on 3.5, rejected on 3.6, and ignored-and-billed on 3.7. Three behaviors in three consecutive minor versions is why this PR trusts the SDK enum and provider errors instead of a model matrix.

Behavior change

reasoning_effort="none" previously kept thinking active (billed, hidden); it now genuinely disables it. Models that cannot run without thinking now raise Google's 400 instead of silently billing. Callers who relied on "hidden but active" thinking should use "auto" or a graded level. Default ("auto") behavior is unchanged.

PR Type

• 🐛 Bug Fix

Relevant issues

Fixes #2 (staged on this fork; re-pointed to the upstream issue number at submission)

Checklist

• I understand the code I am submitting.
• I have added unit tests that prove my fix/feature works
• I have run this code locally and verified it fixes the issue.
• New and existing tests pass locally
• Documentation was updated where necessary (not applicable)
• I have read and followed the contribution guidelines
• AI Usage:
  • No AI was used.
  • AI was used for drafting/refactoring.
  • This is fully AI-generated.

AI Usage Information

• AI Model used: Claude (Fable 5)

• AI Developer Tool used: Claude Code

• Any other info you'd like to share:

• I am an AI Agent filling out this form (check box if true)

[JamMaster1999]

Submitted upstream as mozilla-ai#1294