Candidate B is not accepted for production. Exact source 3e5e094bb17877674f4f3ea027a25fc0bdecdfe7 passes formatting, all 87 selected diagram tests, strict desktop/all-target Clippy and the matching Windows GUI build. The initial independent source review preceded formatting and strengthened clearance/invalidation controls; production geometry policy stayed unchanged. Its original hashes and scope are retained in source-review.md, rather than relabeled as a review of new bytes.

The fresh native reader sees only the complete target scene before locking answers. Initial participants are correct, but the conversation still has plausible foreign Allocation ownership and the pale cue cannot be used confidently at Fit52% or one zoom step. Request/report tracing remains uncertain. This fails the predeclared reliable-association/no- explanation gate. Correct guesses alone do not establish an intuitive workflow. The separately published baseline fresh static review assigns the conversation to that Allocation and its arrowhead, but different readers and static/live conditions prevent a causal effect estimate or human-study claim.

The existing-path 18-world-unit dotted cue changes neither model nor routing. Native candidate and released-baseline SVG exports both succeed. The complete byte comparison finds one nonsemantic cue path; removing exactly that path makes the entire SVG byte-identical to baseline, including every original drawing mark, label and metadata. Both complete 138,661-byte input models stay exact, both desktop processes close normally with exit0 and empty streams, and executable/runtime pins remain unchanged. This is conservation evidence, not repaired reading.

Experimental desktop SHA256: 38abdb21278f85b6d4f946d25535c084aa49eddfac74e99f78caa85ace1faa04. Baseline desktop: d55ecad4b0f75c0c9c26625e5441bef9c9360890b65bc13236c6ebfaf03d631e from source 6e7c4ddaf0775198384a8c4882809ea8e14b2976. Both use runtime f45604a12e8bcd839fd86f667c1be1cc7c4f9da2b3c69ff381ac708f0e1ae7d4 and fixture 9db1015369f450d5810c96a01fddb8a009d57f6aafbda5182703f15ded23d9c8.

The byte-preserved packet includes raw screenshots, locked/continued reading, setup/export reports, input models and actual SVGs, launch/exit receipts, empty streams, scripts, gate logs and source identities. Historical scripts retain absolute paths; adapt to a fresh isolated trial. Do not overwrite prior receipts. B's first launcher stopped before process creation because Windows manifest keys used backslashes; the corrected launcher normalizes keys while preserving the original manifest. Baseline chooser's first typing attempt left its field empty; explicit pointer focus/native path retry succeeds. These tool/harness limitations are not recorded as product defects.

The first SVG checker used a regex-byte membership predicate that did not itself prove baseline cue absence. Full cue-removal byte equality already passed; the final checker adds the direct literal absence assertion, retains the initial checker/receipt, and passes again against unchanged actual exports.

Reproduce code gates from the source checkout with Rust configured:

cargo fmt --all -- --check
cargo test --locked -p oa-desktop -- diagram:: --test-threads=1
cargo clippy --locked -p oa-desktop --all-targets -- -D warnings
cargo build --locked -p oa-desktop --bin arcadia-studio

Build is a development Windows gnullvm executable; no packaged release, Linux, full-suite or full OA/SA acceptance is claimed. Dense rebuild/preview timing, general collision geometry, accessibility/DPI and edited-label persistence were not pursued after the first-reading falsifier. No weight/contrast tuning around this scene is accepted. The next materially different experiment tests measured wrapped labels and placement, with an explicit unresolved authored-format boundary. The permanent Properties pane and narrow canvas remain separate workspace costs; the native chevron expands tools rather than recovering Inspector space.

Independent final evidence review reconciles the complete packet and the failed-reading conclusion.