perf: optimize hot paths with algorithmic improvements

[tnm]

Summary

Three performance optimizations targeting commonly-used code paths:

1. Dependency analyzer: O(n²) → O(1) incoming connections

• Add reverse dependency map built in single O(n) pass
• Lookup incoming connections via dict instead of scanning all nodes
• ~50x speedup for graphs with 200+ modules

2. Line reference fixer: O(n) → O(log n) nearest line lookup

• Use bisect binary search instead of min() with lambda
• Cache sorted line lists per file
• ~100-700x speedup depending on diff size

3. Validator pre-compiled regex patterns

• Move regex compilation to class-level ClassVar
• Combine separate def/class patterns into single regex
• ~1.2x speedup (modest due to Python's internal caching)

Benchmarks

Synthetic (isolated operations)

Optimization	Size	Before	After	Speedup
Dependency graph incoming	200 nodes	31.9ms	0.6ms	52x
Line ref lookup	2000 lines	590ms	0.9ms	696x

Real-world (kit codebase)

Operation	Dataset	Time
generate_llm_context()	335 modules	0.5ms
fix_comment()	2684-line diff	1.1ms

Test plan

• All existing tests pass (46 passed)
• Added 8 new edge case tests for binary search algorithm
• Formatting/linting passes