feat(tooling): bump ruff 0.9.7 -> 0.15.20 (HYBIM-777)

[etserend]

Summary

• Bumps ruff 0.9.7 → 0.15.20 in .pre-commit-config.yaml and pyproject.toml
• Bumps openapi-python-client 0.26.1 → 0.29.0 (0.26.x capped ruff<0.14, blocking the upgrade)

Violations fixed by new rules

Rule	Fix
F401	Remove unused timezone import from utils/__init__.py
PLW1641	Add __hash__ to Message class (defines __eq__ without __hash__)
PLC0415 in src/	Add experiment.py and metric.py to per-file-ignores (circular import pattern)
PLC0415 / RUF043 / E402 in tests/	Add to test per-file-ignores (pytest.importorskip conditional imports)
Multiple rules in examples/	Extend examples/**/*.py per-file-ignores (ANN, ASYNC, S, etc. — examples are docs not prod code)
E402 in create_sample_logs.py	Add # ruff: noqa: E402 (file-level, load_dotenv() must precede imports)

Test plan

• CI green (all Python versions / OS matrix)
• ruff check passes locally ✅
• ruff format applied ✅ (92 files reformatted)
• Test suite passes locally ✅ (1516 passed, pre-existing langchain/openai-agents skips unchanged)

[fercor-cisco]

🤖 This review was generated by the Astra agent (claude-opus-4-8). It may contain mistakes.

Verdict: approve — Well-scoped tooling bump; substantive src changes (StrEnum, datetime.UTC, hash, F401) are behavior-preserving and correctly scoped.

Follow-ups

Suggested follow-up work that could be tracked as Shortcut stories:

• pyproject.toml:248-263: The new examples/**/*.py per-file-ignore list is broad — it silences whole rule families including S (security), B (bugbear), E722 (bare except), and ASYNC. That's a defensible tradeoff for demo code, but wholesale-disabling security/bugbear linting means genuinely buggy example patterns (e.g. broken async, resource leaks) will never surface even when they mislead users copying the examples. Consider narrowing to the specific rule codes actually triggered rather than entire prefixes, so future real issues in examples still get flagged.

[fercor-cisco]

🟡 minor (bug): __hash__ will raise TypeError: unhashable type: 'list' when content is a list (the multimodal list[ContentPart] / IngestMessageContent case that LoggedMessage explicitly supports). This isn't a regression — the class was fully unhashable before (defined __eq__ without __hash__) — but by adding __hash__ you invite callers to put Message in a set/dict key, which will then blow up only for multimodal messages. Consider normalizing list content to a hashable form so the method is total, e.g.:

def __hash__(self) -> int:
    content = tuple(self.content) if isinstance(self.content, list) else self.content
    return hash((content, self.role, self.tool_call_id))

(and if elements aren't hashable, hash their serialized form). At minimum, worth a test covering a list-content Message to document the intended behavior.

Suggested change

	def __hash__(self) -> int:
	return hash((self.content, self.role, self.tool_call_id))
	def __hash__(self) -> int:
	content = tuple(self.content) if isinstance(self.content, list) else self.content
	return hash((content, self.role, self.tool_call_id))

🤖 Generated by Astra

[fercor-cisco]

I'll create follow-up tickets for the 2 notes from the review agent.