Files
ARCADIA-Studio/harness/release-contextual-source-native-v2-plan-20260911.md
T

8.7 KiB
Raw Blame History

Contextual source native v2: remaining acceptance

Codex-9e8r6t /root/behavior_audit, 2026-09-11. Prepared from ca53f574e57381a523e946d5277b4081cff74355, the frozen fixture, unchanged oracle, and v1 evidence. This is a source-derived plan, not new native acceptance by this reviewer. The latest reboot was 13:59:33.500 UTC. Old52's first reopen attempt reached only the file picker before interruption. Root's separate second attempt subsequently completed real Open → SA machine view → Save; the unchanged complete Machine and same-model/original-ID/six-file oracles passed, and the app closed normally. The preserved second-run evidence establishes old52 persistence. Fresh Machine Redo history and acceptance of the new matching GUI/layout remain necessary below.

Keep two independent trials

Use a new .local/native-qa/contextual-source-20260911-v2/ with a newly verified GUI/runtime and private profiles. Preserve v1, its archived evidence, and all original checkpoints. Record final build inputs, executable/DLL hashes, launch/close, client size, each observation, checker exit and immutable checkpoint before advancing. Start at 1100×700 with both side panels. Source-card/caption acceptance requires the new matching GUI; old52 cannot provide it.

  • A — direct Lifeline: copy the complete frozen harness/fixtures/staged-source-authoring/base/ into v2/lifeline/work/.
  • B — fresh Machine history: copy the complete v1 checkpoint-chain-redo/ into v2/machine/work/. Keep v1 checkpoint-lifeline/ and checkpoint-chain/ as its original UUID carriers. Do not seed B from v1 work/ or checkpoint-machine-undo/: those contain the two known detached machine files. Do not combine A's new lifeline UUIDs with B's old chain.

Read-only preparation confirmed the fixture's two published hashes and that v1 Chain Redo has exactly the same four .arcadia paths and hashes as its first commit. This byte comparison does not replace the existing complete core oracle. Before native input, run its baseline/self-check and B's complete Chain check again against the copied input.

Use the unchanged oracle 967D9BA3C6FB3A7EFB410833903EC94556CA2D48A67D5072C8BB160F40A00F9C at .local/staged-source-oracle-20260911/staged-source-authoring-contract.exe. Its matching libunwind is F45604A12E8BCD839FD86F667C1BE1CC7C4F9DA2B3C69FF381AC708F0E1AE7D4; scope PATH to a verified runtime directory plus Windows system directories. The preserved v1 loader failure is not a semantic result. Check nonzero exits even when stdout is empty. Adapt the old wrapper's v1 paths and refuse to overwrite logs/checkpoints; do not alter oracle predictions.

A: cancellation/recovery before direct creation

  1. In Operational assessment, Add first lifeline. For this cancellation-only control, enter parent Retain parent 雨 and child Activity Retain source Δ at the OA root. On initial child open, capture Name, current parent and determined Activity without scrolling; Back/Use must remain reachable. These temporary names will never be submitted.
  2. Before Use or Save, verify the actual completed private recovery checkpoint contains the baseline accepted bundle and both exact raw names, source kind/root, and reserved identities. A ten-second tick alone is not proof that writing completed. Root may stop only this isolated GUI without applying/saving, reopen the same private profile, and use the recovery entry. Recovery must show the raw child in an unsaved local copy, with accepted model unchanged and no inherited Undo history. Compare its authored draft payload with the pre-stop record; envelope timestamps are not draft identity. Original files remain untouched.
  3. With focus in child Name, Escape must dismiss only the child and preserve the parent raw name. Reopening the child must be blank, with its permitted kind/root; unadopted child text is discarded. Enter a valid temporary source, Use, then explicitly Cancel the parent. No source or consumer may be accepted; Undo must report no change. Save the recovered copy only after Cancel, to a separate fresh directory, and require strict same BASE CANCELLED_COPY.
  4. Reopen untouched A's baseline to start clean history. Enter the exact final lifeline Assessment lifeline, create root Activity Assess incident work, and Use. Before Save, record the two independent names, staged source and still-two accepted elements. Click the actual parent Create first, observe selected Lifeline, then Save. This discriminates direct Create from v1's valid Save-adoption route. One Undo/Save and one Redo/Save must use separate immutable checkpoints and the first commit's UUIDs.

Recovery restores draft intent, not the old in-memory history (recovery.rs:Snapshot::restore, app/checkpoints.rs:checkpoint_loaded). Child Escape and parent cancellation follow app.rs:Message::Escape. Save/Save As applies valid drafts; disk equality before Save alone proves persistence only, so pair it with the accepted model/UI observation.

B: complete new Machine history, then reopen

Open the verified fresh Chain copy. Choose SA New → State machine, name Field team operating modes, then Create new owner. At minimum height, Name, SA root and both explicit System/System actor choices must be reachable; select System actor, name Field rescue team, and Use. Check both independent names and the staged New owner label without a false Selected: Missing element. Before Save, accepted state must remain Chain.

Opt into Create the first state view, then click Create and open once and Save. Preserve the first Machine checkpoint. Undo once, Save and check; Redo once in that same live session, Save and check. Only then close normally, relaunch/open through the real picker, Save and compare the final checkpoint. A reboot/reopen before Redo destroys the required history experiment; it cannot be credited as Redo.

The existing oracle expects only root SystemActor, its owned machine, and the requested Field team operating modes · states diagram/root placement. No region or state is implicit. Source names, aliases, defaults, ownership, links, sequence membership and layout stay exactly as predicted in the frozen contract.

Exact checker mapping

All symbols below are entry-file paths. BASE is the frozen fixture; OLD_L1/OLD_C1 are v1 first-commit carriers; A1 and B1 are the newly preserved first commits. Never use Redo/reopen output as its own UUID authority.

self-check BASE
check baseline BASE A_INITIAL
check chain BASE B_INITIAL OLD_L1 OLD_C1
same BASE A_CANCELLED_COPY
check lifeline BASE A1 A1
same BASE A_UNDO A1
check lifeline BASE A_REDO A1
check machine BASE B1 OLD_L1 OLD_C1 B1
same OLD_C1 B_UNDO B1
check machine BASE B_REDO OLD_L1 OLD_C1 B1
check machine BASE B_REOPEN OLD_L1 OLD_C1 B1
same B_REDO B_REOPEN

Counts: A first/Redo has three reachable and physical source files; A Undo has two reachable and exactly three physical files. B first/Redo/reopen has six; B Undo has four reachable and exactly six physical files. Only same … UNDO_FROM permits the explicitly proven retained bytes. Unknown, changed or missing detached files fail; ordinary cancellation/reopen checks remain exact.

Record one pending child at wide size and back to minimum without losing raw input. Observe immediate typing after child Use/Escape and recovery separately: incorrect return focus or scrolling is exposure for peer956's FormShown/workspace work, not permission to change the fixture or claim focus passed. User issue 2's complete canvas workflow remains open.

Overflow regression correction: independent source readback

The real focused3 gate ended 13:57:36 UTC, exit101: 25 passed, one failed at the long-parent fixture's parent_bounds.height > fields.height assertion. Earlier Name, three kinds and first-line assertions passed. The old eight repetitions did not force the claimed overflow; preserve this as a test failure.

Reviewed correction source_draft.rs SHA256 85D543CC288E48B4C13001E4E386B5B54D9B1E349A5AC31E4E307572F2D8FB3C against exact ca53f574 through MCP. Five test-only hunks increase the authored name to 32 repetitions, capture the actual behavior-source-fields content rectangle, improve diagnostics, and additionally require overflowing content containing the entire parent. Real graphics::text::Paragraph, 340×215 host, original paragraph overflow inequality, all visibility/order/footer/raw assertions and production text are unchanged. No material source finding remains. Corrected runtime execution and rebuilt native visibility are still pending; this reviewer ran no compiler, native input or Git command.