[abrichr]

Refs #1331. This PR fixes both causes in code. It does not re-run the openIMIS drift row live (see "Not done").

What was wrong

1. The halt lands on the scroll. The cause is locator resolution, not settle or step order. A closed-loop scroll stops when the next step's anchor resolves. After the relabel, the Save never resolves, so step_014 uses its 2.5x budget and the run halts there. _repairable then refuses it because a scroll is not a click.

I replayed the readiness probe offline on the recorded frames of all three item-50 attempts (~/.cache/openadapt-scoreboard/.../2026-10-19-a/item-050-{1,2,3}), against the launch-4 bundle. On every frame before and after the scroll, the Save resolves by geometry only. Scroll readiness rightly rejects that rung. On the healthy item 49 it resolves by template at the recorded point.

2. RECONCILIATION_REQUIRED is correct, so I didn't relax it. step_002 (Tab) and step_012 (unlabelled picker click) carry risk_review_required. They acted, and nothing proves they wrote nothing. Calling that HALTED_BEFORE_EFFECT would be a false claim of no effect. I found no outcome that was labelled wrong. The existing reviewed-risk binding (openadapt-flow qualify set-risk) clears it. Once a person classifies both steps, the same halt is HALTED_BEFORE_EFFECT and _repairable returns the Save. The test covers this.

Changes

• runtime/replayer.py: when every readiness probe of the next click missed because no direct rung found it, a non-consequential scroll in a linear run hands the miss to that click. The click then accepts only direct rungs (structural, template, OCR, no model). It halts with the usual Could not resolve target for step ... error, delivery_attempted=False and no input. These cases still halt on the scroll as before: a target that resolved but was out of position or failed its identity check, a consequential scroll, and a scroll in a program graph.
• repair/self_repair.py: the not_repairable reason for a run that ended RECONCILIATION_REQUIRED names the unsettled steps. When all of them are unreviewed-risk steps, it points to qualify set-risk.
• benchmark/reliability.py: the handed-off halt is classified as resolution_failed.
• Docs: SELF_REPAIR.md (halt rules), CAPABILITIES.md row, regenerated llms-full.txt. No new commands.

Tests

• New tests/test_relabelled_save_halt.py, a synthetic openIMIS-shaped fixture: Tab and picker (both unreviewed), a scroll up, and a relabelled icon-only Save. It checks three things: the halt is on the Save with no Save input; an unreviewed run stays RECONCILIATION_REQUIRED with an explanatory refusal; a reviewed run is HALTED_BEFORE_EFFECT and repairable. Two of the three fail on main (checked in a detached worktree). The fail-closed one passes on both, as intended.
• tests/test_replayer.py: two exhaustion tests now expect the halt on the click. A new test checks that an identity miss keeps the halt on the scroll.
• Targeted only, because of host load: test_replayer.py, test_program_ir_phase1.py, test_transaction_outcome.py, test_relabelled_save_halt.py, test_capabilities_map.py (213 passed). CI runs the full suite.

Not done

• The openIMIS drift row was not re-run. Its colima VM (openadapt-demo) is stopped. Starting it plus the compose stack under a load average of about 190 would delay the timed measurements, so I left it.
• The re-run needs a person's review first. Before re-running, someone has to classify step_002 and step_012 on the launch-4 bundle (qualify init, then qualify set-risk ... --classification state_changing). In the scoreboard, a scripted reviewer would stand in for that person and should be disclosed in meta.json (demo: CUA to program on openIMIS claims, plus the scoreboard (arms 1-3) #1303).
• Expected result, not measured. With that review in place, I expect item 50 to halt on step_015 and go to self-repair. The Save's identity template has tokens, so frame campaigns apply rather than needs_fault_probes.
• The feat(flow): self-repair a halt with a short bridge of new steps #1325 bridge isn't needed. The relabel is a one-click replacement.

🤖 Generated with Claude Code