[JacobStephens2]

Summary

Check every process Trial's retained Session evidence for the real fix, including Sessions left by a Stop before resume.

Trial completes → scan retained Sessions → pending leak hits hold the Verdict
Day shift dismisses → keep Trial + auditable false positive
Day shift confirms → Void Trial + recalculate paired Verdict
Every decision → reviewer, reason and time

Console forms, CLI decisions and the scoped API share the same locked operation. Experiment results advance to format 6, preserving the updated base's frozen Custom process inputs.

Closes #1880

Evidence

Execution evidence

• Before: PLAYWRIGHT_BROWSERS_PATH=/home/jstephens/.cache/ms-playwright node --test --test-name-pattern='leak decisions retain' console/tests/experiments.test.cjs failed: not ok 1 - leak decisions retain the reviewer and false positives, and confirmation voids the Trial; the CLI returned its usage instead of accepting --leak.
  After: the same command passed: # tests 1, # pass 1, # fail 0. It covers CLI dismissal, Console confirmation, both reviewer identities, retained false positives, immutable decisions, Void accounting, escaping and 390 px layout. Measurement-only hits also hold the Verdict.
• The same browser command also reproduced live refresh erasing the typed reason: live refresh preserves a focused reason, expected The Session saw the real fix, actual ''. After preserving drafts by decision-form identity, it passes with both focused and unfocused fields and verifies the submitted reason.
• Before: python3 /tmp/repro1880-interrupted.py /home/jstephens/repos/thirdshift-issue-1880/target/debug/thirdshift failed with leak_hits: [] despite two retained Sessions containing the exact fix SHA.
  After: the same command exited 0 with two pending merge-commit hits and "verdict": "waiting_on_leaks". The reproduction is retained as tests/experiments/regressions/interrupted_leak.py and runs in the process integration tests.
• cargo test --test experiments process:: -- --test-threads=3: 24 passed, including Built-in and Custom processes, clean evidence, all identifier kinds, abbreviated commits, private-identifier isolation and Stop/resume.
• Result contract test: 1 passed. Python tooling: 41 passed with Python 3.11. cargo check --tests passed.
• CI is green for exact head 27a71cc1: 3,970 Rust tests passed, 91 skipped, plus formatting, all-target clippy, Site and Console checks. Both Linux release builds passed. Rust test execution took 1,172.402 seconds, after 1m07s compilation.
• scripts/affected-tests 15b26e0e2fbf3371773f5232f14456feff719884 passed 68 site checks, 170 Console browser checks, and 1,664 Rust unit tests before stopping on idle::unchanged_requests_reuse_retained_records_but_replacements_are_visible. The same unmodified idle test measures 107,775 bytes on both the Base branch and this head, over its 48 KiB limit; current-head CI passes it. The base comparison used a temporary detached worktree and the same test executable, and the current binary was restored afterwards. Successful CI supplies the remaining Rust validation, as docs/testing.md directs.

Test seams

• Experiment CLI through cached fake Arms and retained Session reads.
• CLI leak decisions and the Console browser flow.
• Console HTTP decisions, token identity and act scope.
• Release script waiting for pending leak decisions.

Visual evidence

Experiment detail with a pending hit, desktop 1440×900:

Same flow, mobile 390×844:

Review

Reviewed against issue-1484, including fresh independent Standards and Spec reviews after reconciliation with 15b26e0e2fbf3371773f5232f14456feff719884.

• Standards: 2 findings, both addressed: glossary vocabulary and an explicit identifier-kind enum.
• Spec: 1 finding, reproduced as written and fixed: stopped Sessions must be checked after resume.
• Fresh reviews of the reconciled code and their coverage follow-ups: 0 additional findings on each axis. Both axes read all 43 changed files, including live refresh, the site test and generated manual. Generated-output exclusions: none. Changed files left unread: none.

Unaddressed findings

Standards

None.

Spec

None.

Security

None.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: one-way

Format 6 adds leak decisions and the waiting_on_leaks Verdict. Experiments waiting on these decisions require a supporting binary; original Trial evidence remains intact.

Blast Radius: Experiments

Built with codex codex-cli 0.166.0 · gpt-6.1-sol · xhigh