perf: Tier 2 optimizations - vector reset and file caching

[tnm]

Summary

Two performance optimizations from Tier 2 (removed module map change after benchmarking showed pathlib overhead made it slower):

1. Vector search collection reset (vector_searcher.py)

• Track reset state to avoid redundant resets on multiple add() calls
• Use delete_collection as primary fast path (avoids count() call)
• Only fall back to expensive ID enumeration when necessary
• ~2-5x faster build_index() restarts

2. Context extractor file caching (context_extractor.py)

• Add mtime-based file content cache shared across methods
• chunk_file_by_lines, chunk_file_by_symbols, extract_context_around_line now share cached file reads
• Simple FIFO eviction when cache exceeds 100 files
• 10-94x faster for repeated reads of same file

Benchmark Results

File Caching (synthetic)

Reads of same file	Uncached	Cached	Speedup
10	0.78ms	0.08ms	10x
50	3.69ms	0.12ms	30x
100	7.49ms	0.08ms	94x

Real-world (kit codebase)

Operation	Result
Context extractor (180 calls on 20 files)	0.88ms avg per call

Removed

• Module map pathlib optimization - Benchmarking showed pathlib object creation overhead (Path(), .parents, str()) made it 10x slower than the simple nested os.path.dirname() loops. The theoretical O(n²) → O(n) improvement was overwhelmed by constant factors.

Test plan

• All existing tests pass
• Formatting/linting passes
• Benchmark verification completed