perf(agents): avoid per-file recomputation and clean up temp files

[DhruvrajSinhZala24]

Overview

This PR resolves repeated recomputation in the Atarashi similarity agents and fixes temporary comment-file leaks noted in Issue #124. Expensive license-corpus preprocessing is performed once per agent instance, and scan paths now consistently clean up temp files.

Key Changes

1. Performance (precomputation / hoisted work)

   • TF-IDF (CosineSim): precomputes the license TF-IDF matrix once; per-file scan only transforms the input and computes vectorized cosine similarity.
   • TF-IDF (ScoreSim): preserves baseline semantics (per-file vocabulary + L2 normalization) while avoiding per-file vectorizer refit via precomputed license-side statistics.
   • WordFrequencySimilarity / DLD / N-gram: precomputes license-side tokenization/frequencies once per agent instance.

2. Resource management (temp-file cleanup)

   • Standardizes AtarashiAgent.cleanup() and ensures it runs via finally in all agents so temp files created by CommentPreprocessor.extract() are removed even on exact matches / early returns / exceptions.

3. Bug fixes

   • N-gram state corruption: avoids destructive mutation of shared license data between scans.

Benchmarks (local, illustrative)

20 .py files from atarashi/ (sorted), 774-license corpus (processedLicenses.csv), single agent instance:

Agent	origin/master	This PR	Speedup
TF-IDF (CosineSim)	4.22s	1.77s	2.39x
TF-IDF (ScoreSim)	3.84s	2.08s	1.85x
WordFrequencySimilarity	3.03s	0.69s	4.42x
DLD	42.69s	39.23s	1.09x

Note: timings vary by machine; speedup grows with larger multi-file scans since one-time initialization is amortized.

Verification

• Ran the CLI smoke suite locally (same coverage as .github/workflows/build-test.yml).
• Compared results vs origin/master on representative files; top matches unchanged for TF-IDF (Cosine/Score), WordFrequency, DLD, and N-gram.

Fixes #124