fix(models): make adapter call-time options explicit

[dvd233]

Closes #63

What changed

• Added a shared validation helper so built-in adapters reject unsupported call-time options with deterministic TypeError messages.
• Added the documented positive-integer max_tokens override to OpenAICompatibleModel, while preserving the provider-default request shape when it is omitted.
• Preserved AnthropicModel's existing max_tokens normalization and made it explicit that max_tokens is the portable option; temperature remains OpenAI-compatible-specific.
• Kept adapter-controlled fields and arbitrary provider SDK parameters out of the call-time kwargs surface.
• Updated the adapter guide and added fully mocked regression coverage for accepted, omitted, invalid, misspelled, and multiple unsupported options.

The compatibility change is intentional: callers that relied on ignored keyword arguments now fail clearly instead of silently doing nothing.

Validation

• python -m pytest -q tests/test_openai_compatible.py tests/test_anthropic.py: 104 passed, 2 skipped
• python -m pytest -q (proxy environment cleared): 206 passed, 6 skipped
• python -m ruff check --select E9,F63,F7,F82 .: passed
• reuse --no-multiprocessing lint: passed
• git diff --check: passed
• Changed-file Ruff, compileall, and pip check: passed

AI assistance

AI assistance was used to inspect the scoped issue, draft the implementation and tests, and review edge cases. I personally reviewed the six-file diff and verified the behavior with mocked provider transports and the validation commands above.

Checklist

• I checked the issue's Development section and open pull requests for linked work.
• This pull request is focused on one change.
• Behavioural changes include tests, or I explained why a test is not
  needed.
• I included the local validation commands and results.
• I updated documentation for user-visible changes.
• I checked ownership before adding or changing SPDX headers.

[knisar]

Thanks for taking this on. I checked the adapter behavior, error handling, documentation, and mocked coverage at 9ffc4f3. I also ran the full suite and tested the result against current main. The two Weave jobs used the PR's older base and hit the test package import issue that is already fixed on main. The merged result passes locally. This is ready to merge.