8.4 KiB
Allocation Arrange policy and pure planner — 2026-09-23
Codex-9e8r6t contribution to issue 29, above Process delivery 6d99aacab8ca1a91c857a2179baa14d80a2901e9.
Pure planner verified; App integration and native acceptance remain open.
This is not a release or full Allocation/multiselect completion claim.
Result and decision
The accepted amendment defines explicit Arrange for one or more work/performer placements. It preserves engineering records and moves one union: semantic descendants of explicitly selected work plus accepted visual descendants of explicitly selected performers. It does not acquire incidental semantic parents or siblings. Ordinary assignment retains its existing per-root responsibility and coherence rules.
The experiment compared strict refusal, visible-selected-only movement and this combined footprint. Strict refusal discarded valid semantics-preserving layouts; visible-only movement omitted hidden/semantic descendants. Actual OA/SA examples support the bounded union policy, subject to complete geometry checks. There is no automatic packing, convergence, semantic reparenting or target unfolding.
Both visible and required diagnostic scenes check all unrelated card pairs. An unselected performer's derived enclosure can grow into an obstacle even when none of the moved cards touches it. Diagnostics unfold affected placements and their required visual ancestors; unrelated stationary folded interiors remain outside that domain. The experiment explicitly distinguishes this policy from universal unfolding. Future expansion can reveal overlap.
At source 9b83f7ca0d,
arrange.rs implements
immutable capture/place/finish, exact original placement identities, one snapped
anchor delta, route translation using actual accepted endpoint owners, combined
projection, every selected header, stale/tamper rejection and an exact final
Project comparison. Motion reuses the semantic index and captured footprint;
it does not clone/validate a Project per pointer position. Pairwise collision
cost remains real; no large-scene latency benchmark is claimed.
The planner's ID map is deliberately not an input capability. App must validate the entire original typed selection before conversion, retaining its exact snapshot/generation and clean-owner lifetime. The new module is not yet connected to App; its temporary dead-code annotation states that boundary. The only change to singleton capture extracts its already-validated fact collector; ordinary singleton policy is unchanged.
Exact verification
All twelve closed runs are preserved in the portable evidence. Every run reports unchanged captured inputs, including all 71 patched Iced files. Recorded compile failures remain failures. Historical HEAD alone does not identify uncommitted captured inputs; named source proofs compare the complete relevant Git inventory and every blob byte with the captured map.
| Gates | Observed result | Exact source |
|---|---|---|
| experiment-1 | 2 OA experiment tests pass | 743 blobs at 3a7923b96d671041cbc747a614cbd0fad6b6b3d3 |
| experiment-2 | Compile failure: imported module resolved desktop-wide tests | Captured capsule retained; no named-commit proof |
| experiment-3 | Four OA/SA experiment families pass, 57 filtered | 744 blobs at 8dbbaf262e3b8d26a32d6c30f0cc5f963fab47a0 |
| planner-document-1 | All 61 planner/Document module tests pass | Same 8dbbaf2 bytes |
| probe-lint-1 / format-1 | Strict scoped Clippy and workspace format pass | Same 8dbbaf2 bytes |
| arrange-planner-1 / -2 | Compile failures: missing explicit module path, then test temporary-array borrow | Both exact capsules retained |
| arrange-planner-3 | 67 tests pass, 2.44s | Earlier captured planner/tests; not labelled as 8dbbaf2 |
| arrange-planner-4 | 70 tests pass, 2.77s | 746 blobs at 9b83f7ca0dcd34056765b9399863a21da72c3a9f |
| arrange-lint-1 / arrange-format-1 | Strict scoped Clippy and workspace format pass | Same 9b83f7c bytes |
The final input-map SHA-256 is
c8192fd84fba63fa576975e9fd44f0fa2d6c8ed2e6b5ada3fd4bee076d36f686.
The 70 include nine new Arrange test families and the existing 61 module tests.
Literal full-Project oracles check engineering records, sizes/folds, other views,
internal and one-ended saved routes. Actual Document execution checks compiled
source, reopen and one exact Undo/Redo. Further cases exercise either selected
anchor, parent/child deduplication, role-only work, unplaced leaves, simultaneous
vacated-space movement, preexisting-overlap preservation versus growth, hidden
collision versus unselected enclosure growth, no-op, stale/tampered results and
all selected headers. These are pure/source checks, not native gesture evidence.
The first experiment's untouched fresh-Document assertion was not atomic refusal evidence; the successor removed that overstatement. Independent review also corrected SA oracle literals before execution: FunctionalExchange auto-ports raise projected work height from OA 94 to SA 132, enclosure height 218 to 256, and the hidden overlap area 13696 to 20116. Production sizing was not changed to fit an oracle.
Independent review and remaining checks
Preserved reviews record
policy criticism, the SA oracle finding, the selection boundary and fresh planner
review. No concrete planner defect was found. Review R1/R2 led to independent
collision-cause discriminators, a singleton-refuses/combined-passes vacated-space
case, and unchanged-overlap versus enlargement controls. The final planner
production bytes are unchanged from the reviewed 67-test capture; only
arrange_tests.rs differs in the 70-test capture.
Deeper nested folded ancestry, explicit endpoint attachment fixtures and broader tamper/non-grid cases remain protective P3 coverage suggestions, not reproduced defects or completed checks. App must retain minimum readable text scale; the pure planner validates world-space header containment and finite positive zoom.
App integration must still prove exact whole typed-set admission, deferred selected-member click collapse, an irreversible drag threshold, one-shot Arrange on both performer/work anchors, dirty/source-failure preservation, stale selection generation, cancellation, frozen geometry and truthful preview/counts. Ordinary work-only group assignment needs its own combined semantic candidate. OA/SA native workflows, actual Save/reopen, Scenario group operations and aggregate deletion remain open. Existing Process routing overlap is not solved here. Linux/WSL verification remains user-paused.
Reproduction and portable records
Verified on Windows x86_64 gnullvm, Rust/Cargo 1.98.1, LLVM-MinGW20260826, one build job and incremental compilation disabled. Initialize the documented project toolchain/DLL environment before invoking Cargo. From the named source:
$env:CARGO_BUILD_JOBS='1'
$env:CARGO_INCREMENTAL='0'
cargo test -p oa-desktop --example allocation_policy_probe --message-format=json -- --test-threads=1
cargo clippy -p oa-desktop --example allocation_policy_probe --tests --message-format=json -- -D warnings
cargo fmt --all -- --check
python -B harness/allocation-arrange-evidence-20260923/checks/verify.py
The focused example imports the actual planner and Document modules. It does not replace desktop/App/native acceptance. The initial full desktop experiment reported 30m42s compile time; the focused successor reported 7.11s. This useful iteration improvement is not a controlled benchmark: target graphs, warm artifacts and machine state differ.
The package retains twelve runs, three compile failures, eight named source proofs, 95 logical files in 85 stored members and seven deduplicated source capsule pairs. Offline verification checks exact inventories, pins, streams, receipts, source capsule contents and proof relationships; it does not execute archived code, independently authenticate logs or repeat Git comparisons. The packager's corruption checks rejected eleven changed relationships and an attempt to overwrite an existing output. Root verified all eleven exact staged evidence files in a fresh export: the portable checks verifier and every review archive/member pin passed without consulting working-copy evidence bytes.