[Zacharon]

A deleted sensitive filename could pass the history scan when its contents were shared with an ordinary filename: git rev-list --objects retained only one name for that blob. Reproduced with the actual scanner on main 22b50db using harmless docs/notes.txt, .env and deletion; the history result was clean. The new duplicate-blob regression fails on that unchanged scanner and passes with this correction.

The history scanner now collects NUL-delimited raw Git diff associations, including roots, merge parents, deletions and renames represented as delete/add. It checks every distinct historical blob/path pair, reads each blob once through one streaming cat-file process, and scans content once per ZIP/text interpretation. Findings and existing source exemptions are applied per path. Credential patterns and archive bounds are unchanged. A related existing basename defect is also corrected: ZIP members are scanned under their own names before the archive display prefix is added, so root-level sensitive members are detected and still reported with their full archive paths.

Added 11 synthetic regressions (20 total): deleted sensitive names, duplicate blobs, unchanged-content renames, ordinary names, deduplicated content scanning, per-path exemptions, ZIP aliases, root-level sensitive members and member/byte bounds, Unicode/space-containing paths, merged side history, shallow-clone reporting and an empty root tree. Documentation now distinguishes default/history coverage and states memory, per-archive, ref and pattern limitations accurately.

Validation at head d080eeaa1ebf2ab47b703b9ca9a0a9709daee7a4 (tree verified equal to tested local tree):

• Linux CPython 3.12.14: scanner 20/20, repository 8/8, no skips.
• Tracked scan: 264 files, no findings. Final-head history scan: 168 unique blobs, shallow=false, no findings.
• git diff --check passed. Only scanner, scanner tests and adoption documentation changed; distribution/site/package bytes are untouched.
• Local validate_candidate.py refused the unsupported platform as intended: it requires Windows CPython 3.14.5.
• Final-head Ubuntu boundary CI: SUCCESS. Completed log confirms 20 tests, tracked 264 files, history 168 unique blobs, shallow=false, zero findings.
• Final-head Windows package CI: SUCCESS. Completed log confirms Windows CPython 3.14.5, candidate archive/journey validation, 86 packaged tests plus 8 repository tests, 210 local links, owner_decision=pending, gold_applied=false and h3_verdict=null. Windows package CI is the historical candidate journey, not a current public.4 or Windows scanner matrix.
• Both CI runs checked out synthetic merge 4beeec8e2697e729d757aef8886a66f4ef05cfd4; its tree e8c16c0ab018a979c3ca5711b08c42c60e0e2ca4 matches the tested PR head exactly.

Draft pending independent review and owner-controlled integration. Implementer verification is not independent H3. This remains a bounded marker/name scanner, not an exhaustive security audit. No merge, release or deployment is requested by this change.