Add feedback tone check (#69)

[SaimM2007]

Summary

Closes #69

Adds a tone check step to PathReview's feedback generation pipeline. Previously, ContentFilter only caught explicitly harmful content (self-harm, slurs, illegal activity) via regex — it had no concept of "constructive vs. discouraging" tone, and it wasn't even called from the review generation flow. This meant harsh-but-not-technically-harmful feedback could reach users unfiltered.

What changed

safety/content_filter.py

• Added ToneChecker, an LLM-based classifier that judges whether a piece of feedback is CONSTRUCTIVE (actionable, specific, encouraging — even if critical) or NEGATIVE (discouraging, vague, dismissive).
• Added ToneCheckResult dataclass (is_constructive: bool, raw_response: str) as the return type.
• Empty/near-empty content short-circuits to constructive without an LLM call, so it can't get stuck looping on blank input.

rag/generator/review_generator.py

• Extracted the original single-call generation logic into _generate_section_once().
• generate_section() now calls _generate_section_once(), runs the result through ToneChecker, and retries generation up to MAX_RETRIES (2) times if the tone check fails.
• If a section still fails after all retries, the last attempt is returned rather than looping forever, with its confidence score capped at 0.3 and a structlog warning logged so the fallback is visible rather than silent.

Tests

• tests/unit/test_content_filter.py — reproduction test showing the original gap (ContentFilter has no tone concept), plus new tests for ToneChecker: constructive feedback, negative feedback, critical-but-actionable feedback (must not be falsely flagged), and empty content.
• tests/unit/test_review_generator_tone_check.py (new) — covers generate_section() passing on the first attempt, succeeding after one regeneration, and falling back correctly after exhausting retries.
• All 8 tests pass.
• Ran the full unit suite (428 tests) before and after this change: 53 pre-existing failures exist on main in unrelated modules (review_service, resume_parser, security, skill_extractor, etc.). This PR introduces 0 new test failures and 0 new lint errors.

Open questions

Tone classification strictness is a judgment call — the current prompt is tuned to distinguish "critical but specific" from "vague/dismissive," but I'd welcome feedback on whether it's too strict or too lenient in practice.