Anthropic: support strict tool use via ToolInfo.options["strict"]

[wuisabel-gif]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor

What is the current behavior? (You can also link to an open issue here)

Closes #4432.

AnthropicAPI.tool_params_for_tool_info hardcodes the ToolParam fields, so there is no way to set strict: true on a tool definition. When thinking is enabled, tool_choice is (correctly) not sent — so there is currently no way to get both reasoning and a schema-guaranteed tool call on Anthropic.

What is the new behavior?

Setting ToolInfo.options["strict"] = True (e.g. via ToolDef(..., options={"strict": True})) now:

• sets strict: true on the serialized tool definition
• strips extended JSON Schema validation keywords (minimum, maximum, pattern, etc.) from the tool's input_schema, which strict schemas reject — same treatment as the existing response_schema path
• adds the structured-outputs-2025-11-13 beta header when any tool in the request is strict

Tools without the option are serialized exactly as before (no strict key, schema untouched).

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

No.

Other information:

Unit tests cover default (non-strict) serialization, strict serialization/keyword stripping, and beta header presence/absence.

🤖 Generated with Claude Code

[epatey]

@wuisabel-gif, thanks for the contribution.

One gap unit tests can't catch: whether the API accepts what we send. We have live "slow" tests for exactly this (gated behind --runapi + ANTHROPIC_API_KEY, so not run in the normal CI — please run them yourself):

pytest tests/model/test_extended_json_schema.py -k anthropic_strict --runapi

I've pushed two test commits:

1. Failing unit test: strict schemas require additionalProperties: false on every object. The top level gets it from the ToolParams default, but nested object params omit it.
2. Live tests: flat strict tools pass (stripping, beta header, and the thinking case all work ✅); nested-object strict tools fail with 400: "For 'object' type, 'additionalProperties' must be explicitly set to false".

The fix should mirror the response_schema path: deep-copy tool.parameters and run set_additional_properties_false() on it when strict.

One design question: options["strict"] is a bare key, but existing provider options are namespaced (options["openai"], options["grok"], options["mistral"], …). Should this be options["anthropic"] = {"strict": True}, or documented as a cross-provider generic? As-is, other providers silently ignore it.