fix: surface upstream provider errors in failure path

[Marius1311]

Summary

PR #80 (a no-op pre-commit autoupdate) is currently red because 5 real-LLM Anthropic tests fail on main. Same 5 failures exist on main since #71 — they're independent of #80. The visible failure (pydantic ValidationError: SimpleOutput.text Field required) is a masked error: the test's SimpleOutput declares text: str with no default, and when the Anthropic call raises an AnthropicError, BaseOutput.default_failure re-validates and produces a ValidationError about the failure object itself. That hides the real Anthropic API error from CI logs, so we can't tell why Anthropic fails.

This PR doesn't disable any tests. It makes the underlying error visible across all providers (per project preference for generic, modular fixes over per-provider patches), so a follow-up can fix the root cause.

• BaseOutput.default_failure → uses cls.model_construct(...) instead of full Pydantic validation. The failure object can never again mask the upstream cause. Production schemas all declare defaults, so behaviour is unchanged.
• New LLMProvider._fail(response_format, reason, *, exc=None) helper centralises logging across OpenAIProvider, GeminiProvider, AnthropicProvider (OpenRouterProvider inherits OpenAI's path). Uses logger.error(..., exc_info=True) when an exception is passed, logger.warning(...) otherwise — replaces the existing inconsistent mix (OpenAI logged str(e) with no traceback; Gemini/Anthropic logged nothing).
• All 11 default_failure call sites in the four providers now route through self._fail.
• Anthropic's "no tool use found" fallback also captures response.stop_reason and content_block_types so logs can distinguish "Claude refused" / "plain text returned" / "empty tool input".

Test plan

• uv run pytest -m "not real_llm_query" — 154 passed locally
• CI green on all hatch matrix environments
• Anthropic real-LLM tests still fail (expected — this PR exposes the root cause, doesn't fix it), and the failure logs now contain a provider-tagged ERROR [AnthropicProvider] ... line with full traceback
• Read those logs and open a follow-up to fix the underlying Anthropic issue (likely candidates: bump default model, dated model alias, tool-schema sanitisation, or retry/backoff)

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 66.66667% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.75%. Comparing base (514e1ab) to head (0d613cb).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/cell_annotator/model/_providers.py	65.00%	7 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #81      +/-   ##
==========================================
- Coverage   80.84%   80.75%   -0.09%     
==========================================
  Files          15       15              
  Lines        1237     1372     +135     
==========================================
+ Hits         1000     1108     +108     
- Misses        237      264      +27     

Files with missing lines	Coverage Δ
src/cell_annotator/_response_formats.py	100.00% <100.00%> (ø)
src/cell_annotator/model/_providers.py	75.95% <65.00%> (+2.32%)	⬆️

... and 3 files with indirect coverage changes

🚀 New features to boost your workflow:

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