Record exact native deletion evidence and resize constraint experiments

This commit is contained in:
2026-09-12 06:52:07 -07:00
parent 24bb24839b
commit e57c390d49
120 changed files with 15376 additions and 2 deletions
@@ -0,0 +1,35 @@
# Native deletion checks at f7d84fa
This is a bounded Windows result, not issue #15 or release acceptance. Actual source was `f7d84faf30cdb5e3695c2dc16ee0a5f2eac91e71`; private GUI binary SHA256 `adea83fd13b68a517bb43067169e492c67ed3c9364ffbb3aa8896d98596a3612`. The [stage manifest](experiments/delete-native-f7-20260912/staged-binary-manifest.json) records precise full-test evidence reuse and strict/GUI passes. The [archive manifest](experiments/delete-native-f7-20260912/manifest.json) identifies frozen baseline, saved checkpoints, screenshots, helpers and checker receipts.
## Method and results
Root operated the native Iced application through Computer Use in a 1442×932 window. The compact fixture contains two OA allocation views of the same activity, its performer and allocation, a scenario participant representing that activity, and an SA state with an Entry function relation. The second file preserves Unicode comments and CRLF bytes. Valid incomplete-model warnings are retained and reported.
The standalone [checker](experiments/delete-native-f7-20260912/oracle.rs) derives each expected deletion from the immutable baseline using actual core `prepare_command(Delete)` and the normal identity-aware source updater before reading actual output. It compares the complete Project and reachable SourceBundle, including unchanged identities, placements, relations, paths, comments and exact UTF-8 bytes. It is independent of desktop dispatch but deliberately relies on core deletion semantics. Corrected checker compilation and disk/memory falsifiers passed; an earlier generator failure from rejecting completeness warnings is retained rather than counted as success. The archived preparation report describes worker-time unexecuted status; the later receipts establish actual root execution.
Every checkpoint below followed an actual native Save. Undo and Redo were single keyboard operations, without source replacement.
| Native operation | Exact saved result |
| --- | --- |
| OA work: plain Delete | Only selected placement `…001f` removed; semantic work, allocation and second view retained |
| Placement Undo / Redo | Complete baseline / first removed checkpoint restored |
| Shift+Delete confirmation Cancel | Complete baseline unchanged |
| Confirm work deletion | Baseline-derived Delete `…000a`, including all-view placements, allocation cascade and represented-lifeline clearing |
| Model deletion Undo / Redo | Complete baseline / deleted checkpoint restored |
| Click SA Entry row, Delete, confirm | Only state relation `…0036` removed; state and referenced function retained |
| Entry relation Undo | Complete baseline restored |
| Click SA state, plain Delete | Only state placement `…003f` removed; state and Entry relation retained |
| State placement Undo | Complete baseline restored |
| Click scenario participant, Delete, confirm | Only lifeline `…0015` removed; represented activity retained |
| Lifeline Undo | Complete baseline restored |
UUIDs share prefix `de1e7e00-0000-4000-8000-00000000`. Full arguments and hashes are in each checkpoint receipt. Screenshots show actual target-specific confirmations and the resulting scenes; source comparisons establish the preserved model beyond visible pixels.
## Confirmed focus defect and limits
The native allocation sequence click work → Delete → Save → Undo → Save → Redo → Save → Undo → Shift+Delete produced no confirmation. Clicking the same work again made Shift+Delete open the correct confirmation. The [failed-focus screenshot](experiments/delete-native-f7-20260912/evidence/03-shift-delete-after-undo-no-dialog.png) and [confirmation after click](experiments/delete-native-f7-20260912/evidence/04-model-delete-confirmation.png) preserve this distinction. The [source review](experiments/delete-native-f7-20260912/focus-review.md) identified revision-wide revocation plus passive Save suspension. Reviewed correction `7bf57db`, integrated as `24bb248`, requires its own executable and native evidence.
One immediate post-Undo screenshot appeared to retain the saved indicator; a fresh observation correctly showed unsaved changes before Save. This did not establish a dirty-state defect.
The final baseline was saved and verified exactly. Root clicked the normal window Close control; the native window and PID 9808 were then absent. Both redirected runtime logs are empty; no process exit code was retained. This trial does not establish native process occurrence deletion, message endpoint clearing, minimum-size accessibility, reopen, full free diagram editing or Linux behavior. The new Gitea UX requirements remain acceptance requirements despite these successful deletion slices.