fix: derive transition backend from capture

[thymikee]

Summary

Confirm broad iOS screen replacements from the actual settle capture and the immutable pre-action presentation.

This prevents a coherent transitional private-AX tree from winning the default quiet window after dismissing a small modal. Ref targets recover the authoritative presentation from their session ref frame when best-effort pre-action evidence is absent or already polluted. Ordinary overlapping mutations and explicit short quiet periods retain their existing latency.

Scope: 5 files in the interaction settle/stable-capture runtime and focused tests; no expansion outside that module group.

Validation

• Regression fixtures were observed red before the fixes for missing backend provenance, delayed first transitions, missing pre-action evidence, and polluted evidence; all are green after the owning baseline/confirmation changes.
• Model-free live iOS Element replay with the packed build: dismissing the Threads notice settled after 1.833s with only the collapsed composer; attachment actions appeared only after tapping Upload.
• Targeted private iOS benchmark reruns with the same packed build: GPT element-25 completed in 71.1s with the attachment menu visibly open, and GPT element-14 completed in 95.6s on the required Team Standup room-information screen. Luna xhigh judged both successful.
• pnpm check:affected --run: all 7 selected gates passed; 248 files / 1,990 related tests passed.
• Docs/skills unchanged: this repairs observation reliability without changing the public command surface or workflow guidance.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	2.31 MB	2.31 MB	+425 B
JS gzip	758.9 kB	759.1 kB	+136 B
npm tarball	881.3 kB	881.4 kB	+145 B
npm unpacked	3.08 MB	3.08 MB	+425 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.7 ms	28.0 ms	+0.3 ms
CLI --help	69.4 ms	69.1 ms	-0.3 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/script-source-bundle.js	+425 B	+136 B

[thymikee]

Reviewed exact head 5362386b: code-clean, with no architecture or correctness finding. The transition baseline now combines immutable pre-action nodes with backend provenance from the first actual settle capture, avoiding mutable/provenance-less session state; reverting the change makes the 1.2s regression accept the transitional tree at the 500ms window. The two-file diff and live iOS settle evidence are sound. Not fully G2G yet: the PR remains draft, most exact-head required CI is pending, and the body should add the required scope disclosure (2 files in stable-capture runtime/test; no expansion).

[thymikee]

Re-reviewed exact 8c6a3a3f: the backend-less iOS fallback is code-coherent and its new regressions are non-vacuous—viewport-root gating preserves broad-transition confirmation while partial projections/ordinary overlap retain the default window. Scope disclosure is now accurate (2 files, no expansion). Not ready yet: the PR remains draft, exact-head CI is still running, and the body records a current-head 791ms transitional attachment path still under investigation. Resolve/classify that live behavior and update validation before readiness.

[thymikee]

Re-reviewed exact f9bb1f62c313997f6d91c049cb836935686e3ed2: implementation and live validation remain code-clean. The only prior blocker was a two-line explanatory comment that grew settle.test.ts past its 2,361-line ratchet; commit f9bb1f62c removes only those lines, restores the file to exactly 2,361 lines, and leaves the focused settle suite 31/31 green. Required pnpm check:affected --run is green (248 files / 1,990 tests). Non-draft, mergeable, and ready for human review; exact-head GitHub CI has restarted, so merge should wait for required checks, especially Coverage.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-18 19:31 UTC