fix: repair unit and mypy failures caused by upstream drift

[njbrake]

Description

Stacked on #951, which has to merge first. Everything here was invisible until the mcp cap let the suite collect again.

1. openai 2.x made cache_write_tokens required on InputTokensDetails. Building one with only cached_tokens raised ValidationError, breaking test_model_args[OPENAI] and test_tool_error_llm_mocked[OPENAI]. Chat completions usage has no cache-write counterpart, so report zero. The model allows extra fields, so this stays compatible with older openai versions.

2. any-llm 1.28 moved gateway providers into a config registry and dropped the generic gateway key. AnyLLM.split_model_provider raises before tinyagent can strip the prefix, so the strip at tinyagent/agent.py (if provider_name == "gateway") is unreachable. That logic lives upstream, so any-agent cannot fix the behavior here. The two gateway: params are xfail with strict=True and raises=UnsupportedProviderError, so when upstream handles the new naming the xfail fails and forces its own removal.

Worth a decision on point 2: is gateway: meant to keep working (fix tinyagent to recognise mzai), or is it retired and the test should go?

A third group, which only becomes visible once #950 lets pre-commit bootstrap and reach mypy:

3. google.adk.tools resolves names through a lazy importlib loader with a TYPE_CHECKING block, so mypy treats them as imported rather than re-exported. Import from the concrete submodules. Runtime behavior is unchanged.
4. any-llm 1.28 widened split_model_provider to return str | LLMProvider, so model._provider.value no longer typechecks.

tests/unit: 314 passed, 8 skipped, 2 xfailed, 0 failed. mypy on 3.13 reports no errors.

PR Type

• Bug Fix

Relevant issues

Depends on #951.

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
• 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.

On the tests box: no new test file here. Point 1 is covered by the two existing tests it unbreaks, and point 2 changes an existing test's expectation.

AI Usage Information

• AI Model used: Claude Opus 5

• AI Developer Tool used: Claude Code

• Any other info you'd like to share: Written under @njbrake's direction. The reasoning and decisions are his.

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/any_agent/frameworks/openai.py	61.87% <ø> (+2.50%)	⬆️
src/any_agent/tools/wrappers.py	92.63% <100.00%> (+0.07%)	⬆️
tests/unit/frameworks/test_google.py	97.43% <100.00%> (ø)
tests/unit/frameworks/test_smolagents.py	100.00% <100.00%> (ø)
tests/unit/frameworks/test_tinyagent.py	98.88% <100.00%> (ø)
tests/unit/tools/test_unit_wrappers.py	96.39% <100.00%> (+0.03%)	⬆️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.