Fix: Crash in deep_scan, validate LLM output, ruff cleanup

[Chocapikk]

Summary

• Fix crash in deep_scan.py: suspects_count used before assignment
• Validate LLM output: confidence field constrained to HIGH/MEDIUM/LOW enum
• Diff truncation marker: LLM now knows when context is incomplete
• Ruff cleanup: 32 issues fixed across 11 files (unused imports, f-strings without placeholders, unused variables)

What's NOT in this PR

Patterns, thresholds, and skip pattern changes from the previous version have been dropped per your feedback. Specifically:

• Thresholds untouched: fingerprint similarity stays at 0.05, silent scan layer2 stays at 20
• Skip patterns untouched: no negative lookaheads added on bump/deprecate/release
• No new msg_words: keyword list stays as-is
• No any()/list comprehension refactors: kept your style

Changes

Bug fixes

• deep_scan.py: suspects_count was referenced before assignment on progress print (line 149), crashes on first batch of 50 commits. Replaced with len(suspects)
• deep_scan.py: removed unused raw_combined variable
• analyze.py: LLM confidence output now validated to HIGH/MEDIUM/LOW, falls back to LOW on garbage
• analyze.py: diff truncation adds [diff truncated] marker so the LLM knows context is incomplete

Ruff (32 fixes)

• Unused imports removed: RATE_LIMIT_DELAY, Path, os, sys, date, defaultdict, README_DAYS_SHOWN, get_advisories_for_date, get_recent_dates, get_all_advisories
• f-strings without placeholders replaced with plain strings
• Unused local variable exc in fetch.py

Test plan

• Run python -m src.deep_scan on a test repo, confirm no crash on progress output
• Run ruff check src/ - should return 0 errors

[Christbowel]

Thanks @Chocapikk.
Cherry-picked the multi-language patterns (23926f6) and regex caching (d3a9930). Kept our existing thresholds and rate limit handling. Great contribution, looking forward to future PRs. Closing in favor of the cherry-picks

[Chocapikk]

Bro tu cherry-pick les patterns et le cache mais tu laisses les bug fixes? 😂 Le fingerprint threshold à 0.05 c'est 5% de similarité, ça matche n'importe quoi. Le skip pattern bouffe les security releases. Et deep_scan.py a une variable undefined qui crash si tu tombes dessus. Le ruff c'est 41 issues dont des unused imports et des f-strings vides.

Fais confiance stp gro 😂😂 tu veux que je split en PRs séparées ou tu reprends le reste?