Files
ARCADIA-Studio/harness/release-diagram-removal-windows-20260912.md
T

6.9 KiB

Diagram removal verification — 2026-09-12

This records source-specific Windows evidence, not a completed release or whole-issue acceptance. Gitea issue #15 owns live work. Linux remains deferred under the separate networking instruction.

Closed runs

Source Command scope Result Identified unchanged inputs
0d6f677329316469b4376ceb45c3ce6dafbb093a cargo test --locked -p oa-desktop --bin arcadia-studio diagram_removal_tests -- --test-threads=1 10 passed 529
dc1506f2e8fd2fb4e92efaba5b1d9425a6129e47 complete desktop binary suite 805 passed, 2 failed 529
28e21a0ae8090f62106fcb264835ab7457a0d299 same command with filter app::diagram 19 passed 530
cb46ef40543bff12634e7fa823a588224d54f3cf complete desktop binary suite compile failed: scenario tool accessor visibility 531
4c387ccd640976ba89dd03ee922a3f63ff92b07c complete desktop binary suite 817 passed, 1 failed 531
42511f0f8ae46f29e11ae18c1030face654c2994 complete desktop binary suite 818 passed, 0 failed/ignored 531
24bb24839b14353c1a2b668f3bac34235d39f7a7 filter app::diagram after native focus correction 22 passed, 1 failed 531
e2c027e54c6be0d051820374f669c52278d90f97 strict all-target desktop Clippy after combined canvas corrections failed: two obsolete Focus event variants 533
40154983cc86ac0a45b062d7381124f262adbb3d complete desktop binary suite after panel-mode cleanup 833 passed, 1 failed 533
83f53c7c6e277e250f06d2ae6bfa6a7988bd2e4b complete desktop binary suite after typed activation evaluator correction 834 passed, 0 failed/ignored 533

All ran through .local/run-allocation-gate.py with .local/toolchain/env.ps1, Rust/Cargo 1.98.1 Windows GNU LLVM, LLVM-MinGW 20260908, serialized compilation. Full command omits the filter. Exact start/end, exit codes, commands, source SHA and file hashes are in the archive manifest and linked receipt files.

The full dc1506f failures were the two legacy raw-Canvas F2 tests assuming private canvas focus. Architecture/state keyboard ownership had moved to the mounted Focusable wrapper. Their rename/identity/placement/camera/history checks remain. New actual App-tree F2 coverage verifies initially unfocused refusal, pointer focus, inline rename, clean Inspector focus, Tab followed by F2 in one event batch and subsequent refusal. This replacement evaluator passes in the 28e21a0 targeted run; a later combined full run is still required.

The independent candidate review found placement-ID selection after Placement properties was rejected against semantic identity. dc1506f normalized admission and retained raw selection separately while explicitly capturing the semantic deletion target. OA/SA mounted coverage verifies the generated menu action, blank-canvas focus, exact placement hiding, semantic confirmation and whole-source Undo. Those cases passed in both applicable runs.

The 28e21a0 run also verifies actual Entry/Do/Exit StateAction and Internal StateTransition row clicks, captured relation confirmation, Cancel, source/model consistency, preservation of reused Function/other state relations, collapsed source/ancestor refusal, and one-step exact Undo. Canvas, App and menu share the renderer's relation-row visibility rules. Row design and evidence boundaries.

The cb46ef4 compile failure was E0624 at the App and menu scenario-tool guards; 4c387cc exposes the existing accessor within the crate. The full 4c387cc run completed on 2026-09-12 at 12:33:28 UTC with all 531 tracked inputs unchanged. All nine derived behavior-removal test functions and the architecture/state removal cases passed, including rendered confirmation buttons, real Inspector click/Tab ownership and F6 reentry. Its sole failure was production_canvas_keyboard_and_delayed_tasks_respect_real_chooser_and_review_owners, which expected a second F6 after deliberately withholding the first focus operation's Apply pass. The pending-focus guard now refuses that intermediate keyboard request. The evaluator must retain its stale-operation challenge while respecting normal native focus sequencing; this run remains a failed full gate.

42511f0 fixes that evaluator's ordering without changing production focus behavior. It explicitly refuses a second F6 while Apply is pending, completes the native synchronous Apply, confirms actual viewport focus, and retains the same operation for all stale-owner replay assertions. The complete suite then passed all 818 tests in 443.23 seconds, ending 2026-09-12 at 12:45:42 UTC. All 531 inputs remained unchanged; manifest SHA256 d2ffe3fc9c06e48dc0b621c2430a94e48c91d910736dfb0216def388edf0360b. See the sequencing review.

Strict lint and GUI follow-up

Clippy at 42511f0 found two cloned single-item test slices. The exact borrowed-slice replacement at 260d773 passed the targeted actual keyboard-arrange test; its subsequent all-target lint exposed a production-unused accessor whose callers were all tests. f7d84fa scopes that accessor with #[cfg(test)]. The archived failed lint receipts remain distinct from the subsequent passes.

At f7d84faf30, cargo clippy --locked -p oa-desktop --all-targets -- -D warnings and cargo build --locked -p oa-desktop --bin arcadia-studio both passed. Each identified all 531 inputs and observed no changes; manifest SHA256 8c35625a7ba22f210219e8f4e0d61605cf09625e752796f1eb48921e4c2edc0c. The full 818-test evidence remains scoped to 42511f0; staging verified that the only intervening tested-code changes were the two borrowed test slices and test-only accessor annotation. No full f7 test run is claimed.

Native f7 testing subsequently found lost canvas keyboard focus across Save/Undo: Shift+Delete required another click. Successful exact saved-model checks do not excuse that interaction defect. Native acceptance and the complete OA/SA goal remain open; current ownership and correction work are recorded in Gitea #15.

The reviewed focus correction at 24bb248 built successfully, then its targeted run ended on 2026-09-12 at 14:04:09 UTC with 22 passes and one failure. The exact allocation Save/Undo/Redo/Undo/Shift+Delete regression, real Inspector ownership during Save, window loss and interrupted/fresh drag cases passed. The persisted-state placement Save/Undo/Shift+Delete sequence did not open its confirmation (diagram_removal_tests.rs:1065). The test remains unchanged as a required regression. All 531 inputs remained fixed, manifest SHA256 53c3c1cb5b0766324ad9338b8101f7317610c182916e18d1debda20b8c8af08d; compile 17m48s and tests 271.47s. This failed gate cannot be reused as native or full-suite acceptance.