[mstykow]

Summary

• Files that repeat a short licence tag on every line (one name.ogg Author / GPL V3+ per line, 300 KB-1 MB) took more than 300 s to scan. Three quadratic steps were responsible; each is now proportional to the work it has to do, with byte-identical output.
  • Sequence matching rescanned the whole remaining query for every single-token hit. The longest-match search now stops once nothing later can beat the best run, and visits only the query positions whose token occurs in the rule range.
  • merge_overlapping_matches removed merged matches with Vec::remove, shifting the tail each time. It now unlinks them in O(1), preserving the existing anchor-removal behaviour.
  • The low-coverage wrapper filter compared every sequence match with every exact match. It now looks only at exact matches near the wrapper.

Issues

• Covers: Copyright detection takes hours on a 17 KB ASCII-art file, and --timeout does not stop it #1446 (follow-up report of 44 slow licence-like files)

How to verify

• Take any of the attached slow-files and run provenant scan --license --compat-mode scancode --json-pp - <dir> on a build from this branch and from main. Output matches on every truncation I compared; the full files no longer time out.
• Measured on a 408 KB prefix of one file, --license only: 28 s on main, 2.2 s here. The largest file (1.04 MB): 45 s after the first fix alone, 11.6 s now. All 44 files complete (--license --copyright, --processes 4, 117 s wall in total).

Follow-up work

• Created or intentionally deferred: these inputs still produce about a million one-token sequence matches that are discarded later, so time grows faster than linearly on the largest files. Pruning them at the source would need a parity analysis, so it is not in this PR.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no blocking issue remains.

Summary

This PR cuts repeated work when scanning long attribution lists:

• Sequence searches skip positions that cannot match the current rule.
• Match merging unlinks removed entries without shifting later matches.
• Low-coverage wrappers check only nearby exact matches.

No new actionable issues were found. The previous test finding is fully addressed, and the profiling prints are gone.

_{Reviews (3) · Last reviewed commit: "perf(license): remove quadratic rescans ..." · Reviewed by Greptile}