[JacobStephens2]

Summary

Each sampled pull request now gets a blind packet before its Scoring window is saved.

sampled pull request
  issue edit history → title/body at merge
  captured PR base → merge commit → complete merged diff
  merge commit → repository files (Git metadata and attribution removed)
  pull request → title/body with attribution lines removed
  CI → check names and conclusions
  private packet → saved window

The sample retains Harness attribution separately. Missing historical issue text fails the Command without advancing its window. Grader sessions remain separate work.

Closes #1025

Evidence

• Before: cargo test --test scoring packets_are_blind_and_keep_the_issue_and_repository_at_merge failed at saved packet directory.
  After: The same test passes for Claude and Codex packets, including removal of commit messages and exclusion of a later fix.

• Before review fixes: the reviewers' exact commands failed with first PR commit missing, packet repository leaked Built with, and blinding rewrote a binary repository blob.
  After: each exact regression passes and stays in tests/scoring.rs:

  • cargo test --test scoring packet_diff_includes_every_commit_in_a_rebased_pull_request -- --exact --nocapture
  • cargo test --test scoring packet_repository_removes_attribution_for_every_harness -- --exact --nocapture
  • cargo test --test scoring packet_blinding_preserves_binary_files_with_valid_utf8_bytes -- --exact --nocapture

• Focused Scoring coverage also checks paginated history, deleted revisions, private storage, source literals, binary files, symlinks, executable bits and exclusion of earlier base changes.

• Before the final fidelity fix: cargo test --test scoring packet_diff_preserves_trailing_whitespace_in_the_last_changed_line -- --exact --nocapture showed the final diff line losing its spaces.
  After: the same test passes; cargo test --test scoring passes all 19 tests (one fixture entry point ignored).

• cargo check and cargo clippy --all-targets -- -D warnings: passed.

• scripts/affected-tests 9ab127320007357c695c5f90593f7417c2312286: site checks passed (43 tests). Broad Rust validation is reused from successful CI on final head 3296899, as docs/testing.md permits. Full Rust tests, formatting, Clippy and both browser jobs passed.

• Timing: /usr/bin/time -p cargo nextest run reached 985.92 seconds wall time, compilation included, before the duplicate local run was stopped after CI passed its head. This is a partial timing, not a completed full-suite benchmark.

Test seams

• thirdshift score / thirdshift scores --json, with fake gh edit history and CI and a real local Git origin; inspect the saved Grader packet and repository snapshot.

Unaddressed findings

Standards

None. The one maintainability finding was addressed by sharing Issue-branch inference.

Spec

None. All three behavioral findings were reproduced with the reviewers' tests, fixed and kept as regressions; no decision was deferred.

Security

None. Guidance-mode review of git diff 9ab127320007357c695c5f90593f7417c2312286...HEAD plus supporting code found no introduced finding with a demonstrable cross-boundary impact: git/gh run via argv with no shell, merge/base commits validated to 40 hex, branch-derived issue numbers are digits-only with strict IssueUrl parsing, the fetch host is fixed to github.com, symlinks are never followed, NUL-bearing blobs are preserved, and packet storage is 0700. No proof-of-concept failed, so no fix and no new regression test. cargo test --test scoring passes 19 tests.

Security review outcome

No unaddressed introduced findings.

Review coverage

Standards and Spec ran independently against fetched issue-978, pinned to 9ab127320007357c695c5f90593f7417c2312286, at e37266e and followed up at 77da798. Both axes read all 13 changed files, including the generated site/docs/index.html hunks. No generated files were excluded, and no changed file was left unread. The parent then closed the follow-up binary finding and the final whitespace preservation check in src/scoring/packet.rs and tests/scoring.rs, retaining both regressions.

Merge Danger

Door: two-way

Rollback stops packet generation; existing sample files remain readable.

Blast Radius: Scoring

Scoring now reads issue history and CI and fetches each sampled merge commit. Failure in packet construction fails that Command before saving its window.

Built with codex · gpt-6.1-sol · xhigh