[Patrick-Erichsen]

Summary

A valid CSV field's closing quote can be mistaken for the start of a removal marker and leave ordinary data incompletely inspected. Prove exact closing offsets against Python's CSV parser before refining the existing loose fallback; every field body and explicit removal instruction remains available.

 declared-marker fallback on complete .csv input
+ bounded field/source mapping
+ exact agreement with csv.reader(strict=True)
+ structural closing quotes cannot open another marker

Evidence

• Before: actual scanner CLI scan <original-file> --no-llm --format json reports partial inspection, 0/1 file and obfuscated_instruction_text for unchanged Solaris products.csv and stacks/threejs.csv.
  After: both report complete inspection, 1/1 file, no exceptions. Each retains its same finding; every issue field except the per-run random finding ID is equal.
• 16 focused controls, 1,164 related producer controls and all 294 native runtime-adoption tests pass. Thirteen retained original inputs preserve every finding signature; only the two CSV marker limits change.
• The complete runtime patch freshly applies to pinned upstream 95a52d8; all 119 source/dependency files match reviewed producer 4b0272f. Independent producer and consumer P0–P3 reviews are complete and scoped-clean.

Merge Danger

Door: two-way

Blast Radius: scanner

Ownership requires complete .csv source, at most 1 MiB and 8,192 fields before canonical parsing, exact field/source agreement, and the existing caller deadline. Malformed, unsupported and over-budget inputs retain the conservative fallback. This is static component proof; full Solaris semantic/prod recovery remains separate.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.