[abrichr]

Stacked on #1714; retarget to main after it merges.

Result. Backend-neutral self-repair (#1709) now runs end to end on native macOS. The test is a native Cocoa test app with a SQLite database, driven through MacOSBackend. The AX tree, screen capture, input and database are live; the repair agent and the approver are scripted.

1. The run stops at a new "Update available" notice before any save.
2. The "Later" click is admitted on element_and_record_check (button 'Later').
3. The probe passes: nothing written, target found, same record, same fields.
4. The replay campaign passes (1/1).
5. The fault campaign passes (1/1): another patient's chart stops before the save.
6. One person approves through the decision portal.
7. The resumed run is VERIFIED, with exactly one new row for the patient.

Under demo, and with no verifier, the click is refused and nothing is written.

That passing run holds the app's caret steady (--steady-caret). With a blinking caret, the save click is refused at the frame lease and nothing is written: #1713.

Changes

• repair/self_repair.py, _pre_save_checks: on a backend with no form to read, OCR may glue a field's words together (REPAIREDRUNFOLLOWUPINTHREEWEEKS). A value of two or more words and at least six letters now also counts when it appears on one line with its words joined (_glued_on_a_line). Single words, and any value too short, keep the word rule. This is a screen pre-check; the record check verifies the saved value.
• benchmark/repair_every_backend/qualify_macos.py:
  • provisions a throwaway approver roster in the work folder, the way the test fixture does, so the scripted answer is signed like a phone answer;
  • adds --steady-caret.
• The benchmark README, the design note's "Qualified so far" section, the results JSON and the changelog.

Tests: _glued_on_a_line unit test; the self-repair, auto-repair and element-identity suites: 80 passed.

This changes a repair check, so it needs its own adversarial review.

[abrichr]

Independent adversarial review

The reviewer was an independent subagent, given the diff and the product's safety rules, read-only. There were three rounds.

Round 1 (9ea78be): REQUEST CHANGES.

• Blocker: the glued same_fields match was a raw substring. "5 mg bid" passed against "25 mg bid", and "John Smith" against "JOHNSMITHSON".
• Non-blockers:
  • the evidence named the parent commit;
  • the blinking-caret result had no file;
  • "one person approved" overstated a scripted approver;
  • the approver roster settings leaked into the app subprocess.

Round 2 (0c3e3f1 and a355989): REQUEST CHANGES. The code fixes were sound:

• whole consecutive OCR words only, with the five negative cases as tests;
• git_sha and git_dirty in each result;
• one result file per mode;
• the scripted approver named as such;
• a scrubbed app environment.

The remaining blocker: one doc sentence said the blinking-caret run was refused at the save, while the clean-tree evidence stopped in the probe.

Round 3 (641f3b9): APPROVE. The README and the doc match macos-ax-steady-caret-20261103.json (resumed, VERIFIED, exactly one row) and macos-ax-20261103.json (no_fix at step_001's identity check, no rows). Both were recorded at c6c07d5 with a clean tree.

Open non-blockers:

• the pixel evidence is from 0c3e3f1; the pixel path doesn't use the macOS changes after it;
• the screen check is word presence on the whole screen, not per field;
• a browser whose DOM read fails falls back to OCR instead of failing closed.