[Jammy2211]

Closes #566. Records the human's ruling on the tier-glance Shadow window. Read with PyAutoLabs/PyAutoMind#529 (merge that one first).

What changes

• AUTONOMY.md — new dated section "The glance Shadow-window ruling — 2026-11-04", after "Merge authority follows Consequence — 2026-10-02":
  • evidence: 69 counted rows, all stage-1 merged-unchanged; spot audit of 8 rows across every risk class → 8 of 8 ok;
  • grant: in-turn glance auto-merge confirmed, narrowed — a Witness: that did not PASS (none, n/a, partial, waived) is a human /prm;
  • not adopted: "public API / default / library return value is judge whatever the tier" — it was conditional on a minded row and the human marked rows 3, 4, 6 ok;
  • stage-2 sub-window for unattended (watcher) merges: counted apart, each of the first 10 needs the different-model adversary leg actually run;
  • ledger accuracy: only organism-made merges count; fixed-model-json (PyAutoFit#1708) stays counted (human did not recall typing that /prm);
  • revert condition: any reverted row demotes glance at once, stage 1 or 2.
• "Unattended glance merges" points at the sub-window; Consumers gains an entry.
• tests/test_autonomy_doctrine.py — 5 tests pin placement, the non-PASS witness rule, the audit outcome, the stage-2 sub-window and the revert condition. Control: all 5 fail against the unedited AUTONOMY.md.

API Changes

None — doctrine prose and doc-contract tests.

Tests

tests/test_autonomy_doctrine.py 19 passed. Full suite: 1489 passed, 1 skipped, 2 failed — both test_prompt_budget.py::test_click_guard_precedes_custom_handler_and_preserves_full_request[50001-*], which fail identically on main (pre-existing, unrelated).

Tier: judge — merge mode: human /prm. Heart YELLOW (workspace-validation timeout in autofit_test MultiStartResumeNaNCounters.py; no release rehearsal for current source) acknowledged by the human at ship.

🤖 Generated with Claude Code