ARCADIA-Studio/harness/scenario-group-removal-evidence-20260925/native-observation-notes.md

Scenario group removal: parent native observations

Source 43340b94a13f192c3b5d3c3112dbe87246790a44; Windows GNU/LLVM desktop/CLI build removal-app-build-1. Observed with Computer Use @oai/sky on 2026-09-25 UTC. Original captures and unsuccessful input attempts remain in the run directories. No production changes occurred during either run.

Run 1

PID 7904, closed normally at 02:09:03.784416 UTC, exit 0, empty stdout/stderr, immutable inputs unchanged. Generated working copy only; original fixture and oracle were not edited.

• Captures 01–17: Home, file picker and baseline Saved A. The initial nonmaximized 1442x816 view had navigation clicks miss their visible targets. Reobservation and activation did not resolve that; maximizing to 1605x872 allowed navigation. File picker UIA focused-element text reported the search box while the visible filename caret and typed path were in the filename field. Enter opened the verified path. Baseline Save matched all 4389 original bytes.
• Captures 18–23: 100% zoom placed the participant headers above the viewport. A scroll attempt without screenshotId failed with a 16x16 bounds error; the repeated screenshot-bound attempt did not pan. Fit restored the full sequence. These attempts do not establish successful panning or readable 100% header navigation.
• Captures 24–26: click Intake participant, right-click first message, Add to selection. Exactly two selected; Properties identifies primary alias a, participant alias intake_lane. This is contextual additive selection, not native Shift-click coverage.
• Captures 27–32: header Remove opens review for 2 selected removals and 10 updated preserved records, 12 matches in two pages. First page shows intake_lane, a, review_flow, b/c/d; second shows e/f/g/h and saved_a/saved_b. Repeated display names have distinct readable aliases. Search @h returns only preserved h; @h_absent shows explicit no-match text. Overall impact stays 2/10. Enter in focused search does not confirm. Footer remains visible.
• Captures 33–34: explicit Cancel preserves two-member selection and exact baseline source after Save.
• Capture 35: Delete after cancel/save did not open review; canvas keyboard focus was not established. The filename says delete-opens-review but the image shows no review. Do not claim native Delete entry success. Mounted Scene tests supply separate input-routing evidence.
• Captures 36–39: selected-message contextual menu exposes one group removal action; it opens the same whole-set review. Explicit Remove applies once. Selection clears, status identifies 2 removed in one change and preserved definitions. Both surviving participant headers and seven unfinished message bodies remain; Saved A states seven unfinished positions retained until endpoints complete. Save matches the independently pinned 3856-byte expected removed file exactly.
• Captures 40–41: one Ctrl+Z followed by Save exactly restores the baseline. Transient selection stays empty; this is expected and distinct from model Undo.
• Captures 42–45: reselect the same participant/message by click/context Add; open review while original removal remains available to Redo.
• Capture attempt 46: UIA Restore led to a minimized-window capture error. Activation and returned-window recovery produced capture47 at1442x816, showing expiry notice and disabled Remove. Native system menu sizing (48–55) produced1102x732. Compact expired review keeps its notice, Cancel and disabled Remove visible; impact body scrolls. No proof of exact client-area dimensions is claimed; these are capture sizes.
• Capture56: Escape dismisses expired review, retaining two selected. Captures57–59: fresh compact review has enabled explicit action, readable aliases and fixed footer; scroll then Next exposes page2 including both saved views. All rows in this fixture remain discoverable. This is not an arbitrarily long-alias or large-scale paging stress test.
• Capture60: explicit Cancel leaves the group unchanged. Capture61: one Ctrl+Y successfully redoes the original removal after resize/expiry/reselection/cancellation. Capture62 Save exactly matches expected removed bytes. Captures63–64 open preserved Saved B; its two participants and seven unfinished messages remain.

Exact snapshots: 01-baseline and 02-cancel and 04-undo equal baseline SHA256 70c32355b82393c03ef87b24862863e0a06c4dcb00cf2c0a84a7cfd2da687a8a (4389 bytes). 03-removed and 05-redo equal removed SHA256 1cb04cbab8428fc672293f29783fffdeecaf6ec6d81d5439501c32e866c8264f (3856 bytes). Every comparison was full bytes without normalization and passed on its first saved-source comparison. No oracle revision was needed.

Run 2

Fresh PID6756, prepared explicitly from run1 snapshots/05-redo.arcadia, same exact build/runtime. Closed normally at02:12:09.320969 UTC, exit0, empty stdout/stderr and unchanged immutable inputs.

Captures01–10 show fresh launch and File Open. The first filename typing attempt was not visible after fresh observation; Alt+N refocused it, second typing became visible, and Enter opened the exact run2 working path. Captures11–14 navigate the persisted Scenario and Saved A: three reusable OA activities still listed, two participants and seven unfinished messages retained. Capture15 Save and snapshot 01-reopened match the removed file exactly. This verifies source persistence, not history or transient-selection persistence.

Bounded judgment

Parent accepts both runs and current CLI fixture-preflight-1 for these bounded workflows, subject to independent evidence review before publication. No native Shift-click, successful focused Delete entry, keyboard-only confirmation, screen-reader completeness, long unbroken alias stress, native Fragment/SA/reuse expansion, real App source-write failure, Linux, default-branch integration or release acceptance is established. Native accessibility tree exposes mostly window chrome. Core/planner and mounted App tests remain separate evidence for cases not exercised here. Operational capture/input recoveries above are retained and do not count as successful actions.