Skip to content

Latest commit

 

History

History
2427 lines (2057 loc) · 155 KB

UX-09-HANDOFF.md

File metadata and controls

2427 lines (2057 loc) · 155 KB

UX-09: Accessible workflows and release gates

First bounded milestone, 2026-09-11, based on a38ea34, merged in #171. The owner selected automated browser journeys and enforced CI gates before remaining UX-07 art work. This starts UX-09; it does not complete its release acceptance.

Launch-scope decision, September 15, 2026: The owner approved desktop-first early access with physical A4/Letter printing included. The roadmap release milestones now separate UX-09A early-access closure, UX-09B broader device/assistive qualification and UX-09C observed usability. Agents lead A, but outstanding owner artwork approval and real-paper qualification remain pre-launch gates. This documentation decision does not pass any unassessed gate or change the existing performance targets. Earlier milestone results below retain their original scope and source revisions.

Digital-art closeout, September 29, 2026: The owner completed the remaining print-key decisions after the revised stairs were shown in map context. Source cdba235 and the retained key/crypt sheets at f503fd9 identify this review; individual decisions and the six passing local stair-render/export case/engine combinations are recorded in the art closeout ledger. Initial-release digital artwork is now owner-approved. Earlier dated art limitations remain historical; A-PRINT, exact-head hosted evidence, reference performance, supported-version/final-candidate scope and release handling are still open. Approval of artwork is not paper/device or participant acceptance.

Implemented contract

playwright.config.mjs uses the existing locked Playwright 1.62.1 test runner. src/test/ux09.spec.mjs imports the existing creation, shell, audience and session journeys. The runner, rather than each journey, owns isolated browser contexts, timeout teardown, failure traces and screenshots. Session popups share that context, so player-window errors are included rather than silently missed. No assertion is weakened to accommodate another engine.

The initial cross-browser run failed in WebKit on ResizeObserver loop completed with undelivered notifications. Player preview was resizing its canvas synchronously while observing the containing viewport. PlayerPreview.tsx now coalesces resize notifications into one animation frame, skips unchanged display/backing sizes, and cancels scheduled work on projection replacement or unmount. Projection changes, zoom and device-pixel-ratio changes still redraw. Browser error filtering is not used.

CI builds production assets and runs four journeys plus the UX-08 export/offline journey in each of three independent engine jobs. It retains failure evidence and results for fourteen days. The aggregate Browser qualification job requires every engine to succeed. Build and test now blocks on lint errors and on increases beyond the recorded 73-warning baseline. Hook warning debt is not declared fixed. The active Protect default branch ruleset (21362094) requires both Build and test and Browser qualification, with strict up-to-date checks. Its existing PR, deletion, force-push and bypass policies were preserved.

The existing UX-08 runner remains separate because it controls worker bytes and actually shuts down its server. It uses the same production build and locked engines. Build failures, missing engines and journey failures are failures, not skipped qualification. Workflow jobs and browser tests have bounded timeouts, no automated retries, and no continue-on-error.

Scope and local results

Production macOS observations use Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5. All four journeys passed in each engine after the player resize fix. The three-engine UX-08 export/offline journey also passed. The initial WebKit failure artifacts are retained separately from the corrective run.

Journey Automated coverage Still outside this milestone
Guided creation and Library Sample preview cancellation, 40-cell sample, new blank project, deterministic seeded previews, independent generated copies, image trace/reload, scene template copy and continue-last-map Full catalog search/delete/migration/failure matrix from UX-01/02
Editor shell Project identity, dimensions, navigation, command availability, focus return, text-input shortcut isolation, recovery dialogs, backup download, 390/768/1440 widths and enlarged text Complete keyboard-only editing, all landscape sizes and physical mobile input
Publication and preview Persisted public/private fields, discovery undo, DOM/accessibility sentinels, visible counts, absent editing controls, shortcut isolation, player downloads and responsive preview Screen-reader task review, comprehensive contrast and nonvisual spatial editing
Session and local display Prepare/start, reveal/move, turn/round, measure/ping, checkpoints, inspect versus publish, blank handshake, stale epoch, disconnect/reconnect, DM/player reload, quota recovery, CAS conflict, end/resume/discard and unchanged authored source Real mobile interruption, OS-level termination and participant routine-play acceptance
UX-08 exports/offline Actual PNG/SVG/JSON, physical DPI and pixel-identical overlaps, bounded 128-cell print page, cancellation, missing artwork, actual cache readiness, delayed writes, multi-window update rejection and stopped-origin editing/export Physical printer measurement, cross-schema upgrade and browser cache eviction recovery

The integrated Ubuntu run at ab86a35 passed 666 Vitest cases across 56 files, including three new resize lifecycle regressions. All twelve browser journeys and three export/offline journeys also passed there. That run's jobs still failed at artifact publication because the upload action excluded the hidden output directory. The workflow now uses a non-hidden qualification-artifacts directory and still fails if artifacts are missing; exact-head CI remains the landing gate. Browser runs use real IndexedDB, canvas and local display transport. Unit mocks are not the basis for the cross-browser claims. The largest UX-08 raster surface remains 8,415,000 pixels, not an F05 interactive performance measurement.

Commands, ports, selectors and artifact locations are documented in DEVELOPMENT.md. CI results and source revisions attached to the pull request are the landing evidence; local macOS results do not establish Linux CI or physical Safari device qualification on their own.

Critical keyboard workflows milestone

Second bounded milestone, 2026-09-11, based on b3a7e94, merged in #172. It adds the fifth journey to the existing required browser jobs without introducing a dependency, weakening a gate or changing project/session schemas.

Surface Change
Notes Native selection button is separate from the editing form and actions. Spaces and newlines remain input, Enter submits the title form, and Escape cancels the draft. Save/cancel focuses and scrolls the edit action fully into view, including Firefox's short-landscape case. Descriptions preserve line breaks and use readable theme text; labels wrap and badges scale with text.
Initiative Native selection no longer swallows input keys. Enter saves and Escape cancels renaming with focus return; ordinary blur still saves. Visible Up/Down buttons supplement Alt+Arrow and drag ordering. Displayed positions map to original indices when stale IDs are omitted. Reorder status is announced.
Modals One focus primitive covers creation, scene templates and existing dialogs. It excludes hidden/disabled/inert/collapsed controls, contains empty/busy dialogs and restores the opener, remaining sheet or editor fallback. Icon picker contents mount only while open, so its trap registers on every opening. Responsive sheets cannot steal focus from an open dialog. Draft Escape does not dismiss the enclosing sheet.
Panel layout Edit/delete/rename/reorder targets are at least 44 CSS pixels. Notes and initiative use the context panel's outer scroller instead of a squeezed legacy inner scroller. Below 501 pixels in viewport height, Close panel scrolls with content rather than covering the focused action.
Run Current round and turn form a named, atomic polite live region. Existing keyboard coordinate controls and session persistence are unchanged. No new private data enters player projections.

The browser journey is deliberately pointer-free after file-picker fixture setup: creation cancellation, note multiline edits, initiative rename/reorder, scene templates, nested modal resize, backup download, recovery focus return, reload, preparation, token movement, next turn, end and resume. It records five viewport sizes at both 100% and 200% interface text, with focus, actual clipping/hit tests, edit-target size and scoped note/initiative text contrast. It does not substitute programmatic focus for keyboard reachability.

macOS WebKit uses its native Option+Tab route to all controls. Firefox's headless shell does not cycle back from the document edge, so the journey can reverse with Shift+Tab. Native select type-ahead is used instead of platform-specific Home/End behavior. These input choices are recorded, not presented as universal browser defaults. Full browser/OS keyboard settings and screen-reader interaction still need human qualification.

Housekeeping consolidates duplicate creation focus logic, includes the new journey in the existing lint scope, updates contributor/user guidance, and corrects #171's stale pending-merge status. The 73-warning hook baseline remains unchanged. Local artifacts stay outside the repository; CI retains them through the existing upload step.

Local candidate evidence

The integrated production run passed all fifteen UX-09 journey/engine cases. After the final label/badge sizing adjustment, the keyboard journey passed again in all engines. The UX-08 export/offline/update journey also passed in all three engines. The full Vitest run passed 683 cases across 58 files, including 17 new keyboard/focus regressions; the targeted nine-file set passed 82 cases. Build and the existing lint gate pass without adding to the 73-warning ceiling. Existing jsdom navigation warnings remain outside this milestone.

Engine Version Keyboard journey Note layout cases Lowest sampled text contrast
Chromium 151.0.7922.34 Passed 10 10.61:1
Firefox 153.0 Passed 10 10.61:1
WebKit 26.5 Passed with macOS Option+Tab 10 10.61:1

The contrast samples are note labels, descriptions and badges plus initiative names/order in the fixture's default UI theme. They are not every theme, state or interface element. Layout evidence checks the full edit-button rectangle against clipping ancestors and its actual hit target, not just viewport bounds. This caught a legacy inner scroller and a sticky close button covering actions in short landscape windows. Independent review also caught the initially closed icon picker's missing trap registration; its open/resize/close lifecycle now has a focused regression.

Artifacts are under the session's files/ux09-release-candidate, files/ux09-final-layout and files/ux08-regression directories. Exact-head GitHub CI remains the landing gate; local macOS results do not establish Linux, physical-device or assistive-technology acceptance.

Dense-map diagnostic milestone

Third bounded milestone, 2026-09-11, based on 54741da. The owner selected F05 performance work, explicitly declined to designate this unusually powerful development Mac as the reference desktop, and then approved bounded closeout with the cursor fix and measured renderer backlog rather than extending into a larger renderer rewrite. Reference-device acceptance remains open.

src/test/denseMapFixture.mjs now supplies F05 v1: 128 x 128 cells, 100 tokens, 200 furnishings, 100 notes, derived room/river geometry, dynamic fog, lights and art enabled. ux09Performance.spec.mjs adds a production import/reload, hover/paint/token/pan diagnostic plus a delayed-draw probe regression to the existing three-engine runner and required browser jobs. Durable edits and single-action undo are assertions; latency remains diagnostic, not a CI performance certification. No dependency, schema, rendering cache or additional production raster allocation was introduced.

The reproduced product defect was full map/art repainting for coordinate-only cursor HUD updates. MapCanvas.tsx now depends on cursor position only when the selected tool actually paints a cursor preview: marker, light, clipboard selection or an in-progress polygon. Paint/drag edits still redraw normally. Five component regressions preserve the HUD/no-redraw contract and the four live preview cases, alongside the existing pointer-cancellation tests.

Local observations

Host: Apple M5 Max, 64 GiB RAM, macOS, 1440 x 900, DPR 1. These are headless desktop observations, not typical-user, physical Safari or mobile results. The fixture uses 32-pixel cells, so the main raster is 4096 x 4096 even when the map is fitted to the viewport. No detail settings were reduced.

The initial frame-only Chromium probe recorded hover p95 values of 251.5, 259.6 and 256.0 ms across three repetitions. Profiling identified drawDitherEdge / native fillRect calls as the largest painting cost, with Folio strokes and furnishing rendering also contributing. Removing the unnecessary hover repaint does not fix those costs during actual edits.

The finalized probe waits for synchronous canvas drawing before its frame opportunity for paint/token edits. An earlier frame-only probe was discarded for edit comparisons because CPU throttling exposed frames occurring before React's delayed drawing effect. A deliberate delayed-draw regression protects against that misleading low-latency result.

The table gives the range of each repetition's p95, not an average or pooled percentile. All three repetitions are retained, including the profiled first Chromium run. Warm-ready figures are navigation-to-driver-observed readiness, including saved-state checks and a frame opportunity, not a shell-ready mark.

Engine / version Hover p95 ms Paint-drag p95 ms Token-drag p95 ms Pan p95 ms Warm-ready seconds
Chromium 151.0.7922.34 5.9-16.8 261.1-269.5 258.8-263.8 12.0-12.7 1.48-2.00
Firefox 153.0 7-13 265-268 261-268 13-14 1.24-1.25
WebKit 26.5 31-33 199-212 201-210 37-39 1.47-1.49

Each repetition retains 24 hover/paint-drag/pan samples, twelve token-drag samples and separate start/commit samples with their maxima. Chromium's unprofiled final repetitions had hover p95 of 8.5 and 5.9 ms; its profiled first repetition was 16.8 ms. Do not turn these into a physical latency, refresh-rate or representative-device claim. The painting bottleneck is still present even on this high-end host.

The initial candidate at 5241fe0 completed all six diagnostic/probe cases. Local evidence is in this session's files/f05-final directory. Earlier f05-baseline, f05-profile and f05-candidate evidence is retained separately; do not mix their different probe revisions into one benchmark population. Reproduction commands, sample semantics and artifact names are in the development guide.

That candidate's 4x CPU-throttled Chromium stress run failed its existing 180-second limit during the third repetition. Its first two completed repetitions recorded paint-drag p95 of 1139.5/1117.2 ms, token-drag p95 of 1174.8/1109.3 ms, hover p95 of 11.7/10.8 ms, and warm-ready times of 5.14/5.05 seconds. These are partial observations, not a completed three-run result or an estimate of a typical device. files/f05-throttled-final retains the explicit incomplete result, two completed repetitions, CPU profile, timeline and timeout trace. The earlier frame-only throttled run also timed out; its misleading paint-drag timings are superseded, not improvement evidence. No timeout, sample count, assertion or roadmap threshold was relaxed.

CI follow-up: independent repetitions

The required Linux WebKit job for 5241fe0 then exposed the same workload partitioning problem without artificial throttling. On its AMD EPYC 7763 runner, the first complete repetition reported paint-drag p95 of 2447 ms, token-drag p95 of 2456 ms and warm-ready time of 4.70 seconds. The test timed out during painting in repetition two because all three repetitions shared one 180-second journey budget. The five existing workflow journeys and the delayed-draw probe passed; this was not an assertion failure in those flows.

Each repetition is now an independent test with fresh browser storage, the same full F05 import, one warm reload, and all of its original samples and assertions. Each retains the runner's three-minute per-test limit; the twenty-minute suite limit, serial execution, zero retries and blocking CI remain unchanged. The complete diagnostic still requires all three repetitions plus the delayed-draw probe. Individual reports identify the repetition and the required total of three, and keep failures explicit. This changes workload isolation, not the recorded per-action latency or a release acceptance target.

The earlier local and failed-CI evidence remains historical; do not relabel it as a successful run of the independently isolated harness. Source revision, test name and repetition metadata distinguish the two arrangements. Linux's multi-second edits reinforce the still-open renderer bottleneck and the owner's decision not to qualify performance using the high-end Mac.

The independent-repetition candidate completed the delayed-draw probe and all three repetitions locally with 4x Chromium CPU throttling. Each repetition took about 1.1 minutes, within its unchanged limit; the complete run took 3.3 minutes. Every original sample and persistence/undo assertion was retained. Paint-drag p95 remained 1100.9-1112.8 ms and token-drag p95 1081.9-1126.4 ms. This is successful collection of slow results, not a renderer improvement. Artifacts are in files/f05-ci-fix-stress; the new head's Linux CI remains the landing gate.

Bounded edge-strip cache milestone

Fourth bounded milestone, 2026-09-11, based on 7130a7f after #174 merged. The owner selected a 16 MiB edge-strip cache, not the broader static/dynamic renderer invalidation redesign. There are no dependency, schema, artwork, export-policy or release-threshold changes.

Renderer and memory contract

EdgeBlendCache in src/utils/edgeBlend.ts retains only dither strips in at most sixteen RGBA canvas pages, each no larger than 512 x 512, per mounted editor: 16 MiB of raw raster storage, with an independent 16,384-entry metadata limit. It allocates pages on demand, never a second full-map image. These bounds do not include browser/GPU copies, object overhead, the existing main canvas, paper texture or other renderer allocations, and are not a total process-memory guarantee.

Edges are identified by cell and direction. Every traversal rechecks derived geometry, semantic neighbor types and the resolved neighbor color. A changed color replaces that strip in place; removed edges are not replayed. Token, fog and selection changes reuse unaffected strips. Scale, cell size, intensity or opacity changes clear the atlas. Project/level, map dimensions, theme, custom-theme definitions, viewing role and unmount also release it; disabling blending or entering print mode releases it rather than retaining unused pages.

Capacity overflow draws directly with the original detail. Admission stops instead of evicting an entire traversal's working set on every frame. High-DPR maps therefore get fewer cached edges, not lower-resolution art. Non-dither styles, oversized strips, non-pixel-aligned transforms and unsupported drawing state keep the direct path. Export/player-projection renderers do not request this editor cache and retain their existing resolution-independent path.

An initial candidate exposed a startup regression, particularly in WebKit: alternating a strip write with an atlas read caused expensive source synchronization/copying. The final implementation populates all missing strips before compositing them in the original order, and avoids clearing fresh slots. A regression asserts that all atlas writes finish before the first read. The superseded files/f05-final and files/f05-stress observations are not the final loading results below.

Art and behavior evidence

The seeds, dot positions, edge order, colors and opacity formula are unchanged. Intermediate RGBA8 compositing is not byte-identical to drawing every individual dot onto the final background. The new three-engine browser test compares 192 cases per engine across cell sizes 8/32/64, DPR 1/1.25/1.5/2, light/dark backgrounds, cold/warm caches, painting/undo, derived rooms/rivers, theme changes and intensity/opacity changes.

An independent full-size, per-edge compositing oracle detects misplaced, clipped or stale atlas strips, allowing two channel levels for device-origin coverage/compositing rounding. The original direct renderer comparison allows at most 8/255 per channel and mean error at most 0.25/255. Measured maxima were 7/255 in Chromium/Firefox and 5/255 in WebKit, with maximum mean error below 0.237/255. These explicit rasterization tolerances replace an initial, unsubstantiated 3/255 test assumption, not any existing release gate.

The bounded F05 allocation case traverses the complete derived fixture with a small destination canvas, separately from the real full-size interaction diagnostic. All engines recorded:

DPR Retained edges Direct overflow per traversal Atlas raw raster
1 12,268 0 16 MiB
2 3,584 8,684 16 MiB
3 1,760 10,508 16 MiB

An unchanged second traversal performs no new strip rasterization or page allocation. Unit/component regressions cover color replacement, custom semantics, removed boundaries, settings/DPR, capacity, metadata bounds, oversized strips, token/fog reuse, project/theme/role changes and disposal. Existing cursor-preview and gesture-cancellation cases remain intact.

Local measurements

Same development host and F05 v1 as above: Apple M5 Max, 64 GiB, 1440 x 900, DPR 1, full detail. All three independent repetitions and every existing sample/persistence/undo assertion were retained. Ranges are repetition p95 values, not pooled percentiles. Chromium repetition one is profiled in both the baseline and candidate. No representative hardware acceptance is claimed.

Engine / version Hover p95 ms Paint-drag p95 ms Token-drag p95 ms Pan p95 ms Warm-ready seconds
Chromium 151.0.7922.34 4.6-15.9 150.2-158.6 148.8-157.5 11.9-12.1 1.49-2.01
Firefox 153.0 13-14 158-192 156-159 13-14 1.24-1.25
WebKit 26.5 30-38 88-93 84-86 36-41 1.47-1.50

The fresh Chromium baseline in this session recorded paint p95 261.9-266.8 ms and token p95 254.4-257.8 ms: approximately 40% lower edit latency. Firefox/WebKit's earlier direct-renderer numbers remain historical comparisons, not fresh same-session baseline runs. The Firefox 192 ms repetition is retained, not discarded. Chromium's first synchronous drawing stack after reload rose from 1.02-1.04 seconds to 1.08-1.09 seconds while populating the bounded cache. Warm readiness includes polling overhead and one 2.01-second result; do not claim a loading-target pass.

The complete final 4x Chromium CPU stress run took 2.3 minutes, with each repetition inside its unchanged three-minute limit. Paint p95 was 632.4-653.1 ms, token p95 615.0-637.2 ms, and warm readiness 5.09-5.56 seconds. These are still slow results, not typical-device estimates or mobile acceptance.

The integrated pre-batching candidate passed all 33 browser cases and all three UX-08 export/offline journeys. After correcting atlas population, all 18 targeted renderer/F05 cases and four Chromium stress cases passed again, along with 84 targeted unit/component cases. Evidence is under this session's files/f05-baseline, files/f05-batched-final, files/f05-batched-stress and files/ux08-regression. Raw reports identify the base revision and dirty working tree; the PR identifies the committed patch. Exact-head CI remains the landing gate, with no skipped checks, relaxed samples or timeout changes.

CI follow-up: small-map atlas surfaces

The first Linux run at e51d702 passed the build, Chromium and Firefox jobs, and ten of eleven WebKit browser cases. The pixel comparison failed only for 8-pixel cells: cached versus isolated-edge differences reached 55/255, while all 32/64-pixel cases remained inside the existing 2/255 oracle bound. This is a rendering discrepancy, not a reason to loosen that bound.

Atlas page dimensions now also stop at the map's physical backing dimensions, so a small map does not force its edge rasterization onto a larger 512-pixel surface. Canvas raster backends can differ with surface size. This is a shared renderer change, with no browser/OS branch. The sixteen-page and 16 MiB limits, the full-size F05 atlas layout, and every existing pixel assertion are unchanged. A new regression checks rectangular small-map pages and disposal on resize. The custom-semantic cache fixture retains both edge assertions using an empty second row so the strips and their gutters fit the smaller surface.

All six renderer browser cases passed locally after this correction, alongside 41 targeted unit/component cases and the existing build/lint gates. Linux CI at c08e327 reduced the worst oracle difference from 55/255 to 5/255, but still failed the unchanged 2/255 bound for small maps. Failure artifacts remain in files/ci-webkit-failure; local corrective evidence is in files/ci-small-atlas-fix.

CI follow-up: device-pixel compositing

The remaining Linux WebKit discrepancy is not hidden with a larger tolerance. Cached strips now copy directly in physical pixel coordinates with equal source and destination extents, instead of dividing their dimensions by DPR and scaling them back through the destination transform. The copy preserves the integer viewport translation and restores the caller's drawing state. It also retains the caller's image-smoothing setting rather than forcing a different sampling path from the independent oracle.

The added regression requires integer source/destination rectangles, equal pixel extents, retained translation and balanced context restoration. All 42 targeted unit/component cases and nine local browser cases passed, covering the unchanged pixel matrix and cache bounds plus one full F05 repetition per engine. That repetition's paint/token p95 values were 156.1/148.6 ms in Chromium, 162/158 ms in Firefox and 90/93 ms in WebKit. This is a targeted regression observation, not a replacement three-repetition benchmark. The build and existing lint gate pass; Linux confirmation remains pending the new head's CI. Evidence is in files/ci-device-pixel-fix and files/ci-webkit-small-atlas. No samples, thresholds or timeouts changed.

CI follow-up: bounded pixel diagnostics

Linux run 34700570563 at f9d063e retained the same 192 measurements as c08e327, including 54 small-map oracle failures (maximum 5/255). Browser qualification is a cascade of the WebKit journey failure, not a separate cause. The renderer is unchanged in this diagnostic follow-up.

After the original comparisons, a failure in any engine attaches edge-blend-diagnostics JSON for two 8px cases at different DPRs and one matching 32px control. At most four edges per case expose actual atlas/source rectangles, context attributes, exact vector commands, worst-pixel crops, transparent alpha/premultiplied deltas, same-surface translation replays, and full-source versus cropped-source compositing. Source reads occur after diagnostic blits. The original matrix, inputs, assertions, F05 harness and CI timeouts are unchanged. For local validation only, set QA_EDGE_BLEND_DIAGNOSTICS=1 alongside QA_OUTPUT when running npm run test:browser:run -- src/test/ux09EdgeBlend.spec.mjs. This forces evidence collection, never changes the pass/fail bounds, and is not Linux reproduction or a claim that CI is fixed.

The forced local pass completed all six renderer/allocation cases across the three engines. All selected diagnostic replays matched their original case's isolated maximum, direct maximum and direct mean. Focused lint and strict type-checking of the test-only harness also passed.

Run 34701738152 reproduced the 54 Linux WebKit failures and showed identical visible source pixels but source-rectangle-dependent compositing. The follow-up adds a test-only whole-atlas copy, shifted by destination minus source offset and clipped to the destination strip. The same three cases/four inspected edges report per-edge deltas against the cropped atlas and full-map oracle, plus all-edge composed deltas and a small worst-pixel crop. composedReplay.atlasVsCached checks that the source-subrectangle replay reproduces the measured cached image. All source consumers finish before diagnostic source readbacks. This alternative is evidence only, not a replacement gate or a production fix; Ubuntu must still establish whether it removes the rounding difference.

The forced Mac comparison passed all six browser/allocation tests, with all 192 original measurements per engine identical to the pre-change baseline. Both atlas replays matched the cached image exactly in all nine selected cases. Diagnostics took 80-130 ms; focused lint and strict harness type-checking passed.

CI follow-up: clipped whole-atlas compositing

Ubuntu run 34702635983 at b110564 confirms the same PR-caused WebKit pixel failure: 54 small-map cases exceed the isolated 2/255 bound, with maximum 5/255. Direct-renderer maximum 6/255 and mean maximum 0.217/255 still pass. Browser qualification fails only because it aggregates that job. Build/test, Chromium and Firefox pass; all ten other WebKit cases pass. No transient or pre-existing check failure was found.

The three diagnostic replays exactly reproduce the original measurements, and every cropped all-edge replay matches its measured cached image. For the two failing selected cases, source pixels are identical but cropped source copies introduce per-edge compositing differences. Clipped whole-atlas copies eliminate those per-edge differences, reducing the composed isolated maxima from 5 to 1 (8px/DPR 1) and 4 to 1 (8px/DPR 1.5). The 32px control is unchanged at 1. This supports changing the copy operation, not the rasterization, art, oracle or tolerances. It is bounded evidence for these cases, not yet a passing Linux production-head matrix.

Production now clips a whole-page copy to the physical destination strip and shifts it by destination minus source offset. Each admitted entry retains one rectangle-only Path2D, reused across hits/color changes and released with the entry. This stays within the existing 16,384-entry bound, adds no raster pages, and preserves the caller's current path as well as drawing/clip state. There is no browser-specific branch. Diagnostics understand the whole-page call and report clippedAtlasVsCached as its replay-fidelity check.

Local validation passes 42 targeted unit/component tests, strict harness type-checking, build, and lint with the same five existing warnings. Fifteen three-engine browser cases cover all 192 unchanged pixel measurements per engine, eight physical-placement/clip/path/state scenarios per engine, unchanged F05 allocation limits, delayed-draw probes, and one complete F05 repetition per engine. Isolated maxima are 2/2/0 for Chromium/Firefox/WebKit; direct maxima are 7/7/5, with mean maxima below 0.237. All nine selected diagnostic replays are faithful. DPR 1/2/3 retained-edge/overflow counts and 16 MiB raster limits are unchanged, without warm traversal raster/path churn.

Focused before/after observations on the same M5 Max/64 GiB development host retain every interaction and persistence/undo assertion. Paint/token p95 ms: Chromium 156.2/151.2 to 159.6/155.2; Firefox 161/159 to 159/149; WebKit 102/106 to 98/90. This single paired repetition indicates the cache benefit remains, not statistical equivalence or reference-device acceptance. Candidate warm-ready times are 2.03/1.33/1.58 seconds; release targets remain open. Evidence is in logs/ci-clipped-ubuntu-34702635983, logs/ci-clipped-copy-baseline and logs/ci-clipped-copy-candidate/validation.json in the follow-up worktree. The exact production-head Ubuntu checks remain pending after push; no assertions, F05 samples, CI settings or timeouts changed.

Bounded Folio floor-path milestone

September 14, 2026, based on e6c0b11. The unchanged F05 Chromium profile identified repeated Folio tile drawing as the largest remaining art cost: approximately 1,977 ms inclusive across the profiled interaction sequence, including 1,821 ms in shape drawing and its native Canvas calls.

src/themes/folio-v1/tileCache.ts now reuses only the four variants of flagstone, worn wood and earth floors in MapCanvas. Walls, water, symbols, other themes and all export renderers keep their direct drawing paths. The versioned artwork, source fingerprints, map schema, fog policy, derived geometry, furnishings, lighting and gesture/undo implementation are unchanged.

Admission, memory and fidelity

Each editor owns at most twelve variants and 56 native Path2D objects. The cache retains immutable line/circle paths and rectangle commands. It creates no canvases, bitmaps, image copies or additional raster pages. The existing edge cache's 16 MiB raw raster ceiling is unchanged; native path storage is separately bounded by object count, not an asserted byte estimate.

Floors are drawn into the original destination context, preserving primitive order, materials, contours, resolution and caller compositing. Size changes refresh the detail tier. DPR and transforms are applied at drawing time without raster resampling or a high-resolution cutoff. Project/level changes, dimensions, themes, audience, print mode and unmount release retained paths. Token movement, fog, painting and undo reuse artwork, never cached map state.

Rejected bitmap approaches and Linux investigation: CI run 34837437995 at 0ce50af exposed Linux WebKit floor-sprite differences up to 14/255. Uniform backing surfaces in 34690f7 did not solve the pixel mismatch and introduced a performance regression in run 34838899112: all three F05 repetitions exceeded their unchanged 180-second limits. Traces show 3.6-3.9-second drawing waits, not missing commit events.

The owner explicitly approved a focused Linux investigation after that bounded pass was paused. Diagnostics in run 34840840359 vary backing size, willReadFrequently, source origin and whole/clipped/cropped copies. Changing copy mode or normalizing geometry does not remove the differences. Rasterization changes with surface size and readback hints; drawing at the original origin on an equally sized surface reproduces the reference. That does not justify allocating full-map buffers for every tile variant. Both sprite implementations are therefore replaced, not conditionally shipped. There is no browser/OS branch, no main-canvas backend change and no relaxed image tolerance. The reusable native-path approach avoids cross-surface rasterization entirely.

The new blocking browser matrix compares direct and cached rendering for 210 combinations per engine: 8/32/64-pixel cells, DPR 1/1.25/1.5/2/3, transparent and paper backgrounds, cold/warm traversal, paint/undo, materials, derived room/river geometry and a dense floor grid. Alpha must match exactly; RGB allows one 8-bit rounding step, with mean error at most 0.01/255. Local native-path maxima are 0/0/0 for Chromium/Firefox/WebKit, with zero alpha or mean difference. Eight additional scenarios per engine cover physical translation, clipping, caller state and cold/warm path reuse.

Comparable local observations

Both baseline and final runs use the unchanged F05 diagnostic, three independent repetitions per engine, identical samples and persistence/undo assertions, 1440 x 900, DPR 1, and no CPU throttling. Host: Apple M5 Max, 64 GiB, macOS kernel 25.6.0. Browsers: Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5. These are ranges of repetition p95 values, not pooled statistics or representative-device acceptance.

Engine Paint p95 before / after (ms) Token p95 before / after (ms) Warm-ready before / after (s)
Chromium 153.2-161.7 / 150.6-151.8 151.8-157.4 / 143.8-157.0 1.49-1.54 / 1.49-2.00
Firefox 158-193 / 140-143 164-204 / 138-144 1.25-1.27 / 1.24-1.25
WebKit 91-99 / 77-79 84-98 / 76-86 1.49-1.54 / 1.47-1.53

Chromium repetition one is profiled in both runs; its paint/token p95 changes from 157.3/157.4 to 150.6/147.2 ms. The two unprofiled repetitions change from 161.7/151.8 and 153.2/153.1 to 151.6/143.8 and 151.8/157.0 ms. Chromium's improvement is modest and token p95 is mixed; its last token repetition is slightly slower. Firefox and WebKit paint p95 are consistently lower in this local sample. This is not a universal speedup or a 100 ms desktop-target pass. Earlier approximately 20% Chromium figures described rejected raster prototypes and must not be attributed to the native-path change.

Chromium's warm-ready proxy remains variable, including one repetition at 2.00 seconds. It includes browser-driver assertions as well as drawing. No startup improvement or loading acceptance is claimed.

The current native-path candidate passes 82 focused unit/component cases, strict harness type-checking, lint/build, nine floor browser cases and all twelve F05 cases locally. Raster-specific allocation tests have been replaced with path-count, no-canvas-allocation, high-resolution and caller-alpha coverage. The original pixel matrix, clip/state scenarios, F05 samples, timing targets, retries and timeouts remain unchanged. WebKit runs the pixel matrix as an early CI preflight; failure blocks the job before expensive full-suite execution. Passing it still requires the full unchanged browser and UX-08 suites.

Current artifacts: files/perf-baseline, files/floor-vector-final and files/perf-vector-final. Rejected approaches and diagnostics remain in files/ci-webkit, files/ci-webkit-surface, files/ci-floor-diagnostics, files/floor-surface-fix and files/perf-surface-fix. Exact-head remote CI remains the landing gate. UX-09 and human/device release qualification remain open.

Next bounded performance work

Re-profile the remaining non-floor tile, furnishing, edge and lighting redraws before choosing another bounded optimization. Investigate the Chromium warm-readiness interval separately from first-draw time. A static/dynamic-layer and dirty-region redesign is a separate, larger scope requiring explicit agreement on invalidation and memory budgets. Preserve fog, derived geometry, cursor previews, undo and export behavior; do not silently reduce detail or add unbounded full-map bitmaps. The 100 ms desktop, 150 ms mobile and two-second warm-launch targets remain unchanged and unaccepted. CPU throttling is only a stress probe; reference hardware, physical input-to-paint tracing, mobile/detail policy, memory pressure, cold/offline loading and F05 high-resolution export remain open.

Hook and performance PR cleanup

Bounded cleanup, 2026-09-12, based on 9dd4b10 after #175. The five remaining hook dependency warnings are resolved and npm run lint now permits zero warnings. Selection callbacks track their current project/scope, the canvas selection key listener uses its current owner, draft cancellation/completion declare the stable paint-cell ref, and pointer movement drops an unused zoom dependency. No lint rule or browser gate is disabled.

Older performance PRs are reconciled in this cleanup rather than merged onto newer canvas code unchanged:

PR Reconciled behavior
#155 Preserve cursor-position identity within a tile. #174 already avoids ordinary cursor-only art redraws, but marker, light, clipboard and polygon previews still benefit from this additional guard. Crossing a tile boundary continues updating the preview and HUD.
#164 Remove full-map JSON serialization from draft completion and skip identity-no-op stages. Unlike the original proposal, compare changed branches through structural sharing so dragging a token, stamp or river back to its origin still creates no commit. Unchanged map branches and tile rows are not traversed.

Regressions cover all four preview tools, replacement selection callbacks, serialization-free completion, untouched branch/row access, identity no-ops, repeated completion, changed object keys and return-to-origin moves. Existing gesture interruption and single-commit behavior remain intact. This is not a new renderer architecture or a claim that F05 release targets are met.

Local candidate evidence: 72 cases in eight targeted Vitest files, the zero-warning lint gate and production build pass. Twelve browser cases cover editor navigation, keyboard workflows, the delayed-draw probe and one complete F05 repetition per engine in Chromium, Firefox and WebKit. This focused run is not a three-repetition benchmark or full release qualification. Artifacts are in the cleanup session's files/cleanup-browser directory. Exact-head required CI remains the landing gate.

Library and recovery regression milestone

September 15, 2026, based on c72371c after ART-06/08 merged in #181. The owner selected a bounded, low-input follow-up: promote selected Library and recovery invariants from the standalone development-server qualification into the existing required production-browser jobs.

src/test/ux09Library.browser.mjs registers nine independent cases through ux09.spec.mjs. The existing runner owns their pages, contexts, teardown, traces, screenshots, result attachments and all-tab error capture. The creation journey's native IndexedDB reader is reused for durable-state assertions. No application API, storage schema, artwork, rendering cache, dependency, retry policy or CI timeout changes.

Cases Production assertions
Rich duplication Distinct local identity/revision, complete two-level content and asset equality, portable provenance, copied tags, no copied previous-save/checkpoint records, exact source/checkpoint preservation after copy edits, real private JSON download
Archive and Trash Cancel writes nothing; reload and restore preserve identity/content/tags/recovery copies; no Open control for non-active records; confirmed permanent deletion removes only the target's current/previous/checkpoint records
Competing Library tabs Stale metadata and duplication fail visibly without writing or leaving an orphan copy; refresh shows the winning title
Four stale-editor variants Rename, archive, Trash and deletion in a second tab invalidate the old editor revision; stale saves cannot overwrite or resurrect the durable project; latest unsaved work remains downloadable
Failed duplicate opening A native read fault after commit leaves exactly one durable copy in the refreshed Library; reload and retry open that same identity
Failed save Native quota faults roll back the complete transaction, including the previous-save slot; repeated retries remain failed; backup retains newer edits; successful retry, reload and Library reopen preserve the latest intended work

Fixtures enter through the public import UI; retained predecessors/checkpoints are created through the actual recovery UI. Native storage reads observe committed state rather than replacing repository functions. The two fault cases count page-scoped native get/put failures and restore the methods in finally. This is injected transaction failure, not physical disk exhaustion, OS interruption or natural storage eviction.

The first Chromium pass exposed two new-harness mistakes: an unscoped saved status locator matched both header and dialog, and stale duplication used the metadata-conflict message rather than its own source-conflict message. Those selectors were corrected without changing product behavior or assertions. The cross-engine pass also confirmed that restoring the unavailable current project remounts the Library in Recent maps. The journey now asserts committed active status and explicitly reselects Archive/Trash before checking absence, rather than mistaking the restored Recent card for a failed restoration.

The final local production run on the merged ART-06/08 baseline passed all 27 new case/engine combinations plus the three existing guided-creation cases that share the storage reader: 30 cases in 2.4 minutes. The zero-warning lint gate and production build also pass. Evidence is retained in the session's files/library-final, with source metadata identifying c72371c plus the working-tree patch. Earlier failed harness evidence remains in files/library-first and files/library-three-engine; it is not passing qualification. Exact-head required Linux CI remains the landing gate.

The older standalone suite remains available. This milestone does not promote its entire native repository matrix, failed-startup recovery races, legacy migration/deletion tombstones, thumbnail failures, or Library-specific keyboard/reflow checks. It does not qualify human/device, cross-schema update, physical printing, or performance acceptance. Commands and artifact handling are in Development.

A-DATA production preservation milestone

September 15, 2026. Implementation and browser cases: ca0dedb5f5f0fc4e65fca68d53695cc8649be59d, based on the owner-approved release ledger at 62e0b95. Responsible reviewer: Copilot, using source inspection, production-browser behavior and native storage assertions. Disposition: passed for the bounded cases below on the recorded local engines; exact-head required Linux CI remains the landing gate, not a result inferred from local execution.

ux09Preservation.browser.mjs promotes ten independent cases from the older Library/foundations scenarios into ux09.spec.mjs. Existing required CI automatically runs them on every configured engine with the same runner, three-minute case timeout, twenty-minute suite budget, no retries and all-tab error capture.

Cases Production assertions
Concurrent migration Two startup tabs obtain one local identity and revision; IndexedDB takes precedence; rich project data and both original sources remain exact; repeated root startup is idempotent
Bare-map migration Whitespace-bearing localStorage bytes, embedded custom assets and unknown fields survive migration, backup download and reload
Migration abort/retry Aborting after both adds are queued exposes neither project nor marker; the original remains downloadable; Retry restore commits once
Two unsupported legacy sources IndexedDB and localStorage each retain their original through preview cancellation, recovery, original download from Recovery copies and reload
Deletion tombstone A native abort while writing the tombstone rolls back current/previous/checkpoint deletion; retry commits the tombstone; detached in-memory backup stays usable and Library/bare-root reload never resurrect retained sources
Failed-startup quota recovery Repeated checkpoint quota failures preserve identity and every prior record; retry commits the recovered project with exactly one original checkpoint and survives reload
Competing recovery One tab holds a completed native read while another recovers the same source; creation/editing remain unavailable, the stale restore cannot overwrite newer work or add a checkpoint, and original download plus Retry restore remain usable
Two unavailable-source variants Missing and corrupt startup sources survive an alternative disappearing after catalog enumeration; no blank editor or writes appear, and a healthy alternative opens and reloads

The race revealed a product defect: recoverFailedProject turned a clean failed-startup CAS rejection into ordinary editor conflict state. That exposed the default blank map and removed Retry restore even though no project had loaded. The coordinator now stays in failed-startup state with an explicit retry message when no newer in-memory work arrived. Unexpected pending work still enters conflict and stops automatic saving, preserving the existing in-memory recovery contract. Two coordinator regressions cover these distinct outcomes; existing hook coverage retains the creation-during-recovery guard. No serialized data, migration policy, artwork or rendering behavior changes.

The final run at ca0dedb passed 60 case/engine combinations in 209.8 seconds: 30 new preservation cases, 27 existing Library cases and three guided-creation cases sharing their storage reader. Environment: macOS 26.6.2, headless locked Playwright engines, 1440 x 900, DPR 1, production /Dungeon-Mapper/ assets.

Engine Recorded browser version New A-DATA cases Related regressions
Chromium 151.0.7922.34 10 passed 10 passed
Firefox 153.0 10 passed 10 passed
WebKit 26.5 10 passed 10 passed

Evidence is retained in session 24484730-7e2b-45b8-a5e8-d25ad816604c, files/data-final/browser-report.json, its HTML report and per-case journey-results/page-errors attachments. Every result records the full implementation revision above. The first two runs retain harness failures and the pre-fix blank-editor reproduction in files/data-first and files/data-second; those are not passing evidence. files/data-three-engine is the earlier successful patched-tree run. The build, zero-warning lint and 68 targeted coordinator/hook/recovery component tests also passed.

This closes the bounded A-DATA production coverage gap, not all historical repository probes or every failure mode of physical storage. The same-tab unexpected-pending branch is unit-level evidence, not fabricated UI activity through disabled controls. Cross-version updates, release publication, performance, art approval, physical printing and broader device/participant acceptance remain open. A-EDIT's subsequent milestone is recorded below.

A-EDIT production editing milestone

September 16, 2026. Implementation and browser cases: 4f3fad334852da61af3af2f8e48815aa4e1f6686, based on merged A-DATA 72f32e1. Responsible reviewer: Copilot, using source inspection, production browser journeys, native storage observations and focused screenshot review. Disposition: passed for the bounded local cases below. Required exact-head Linux CI remains the landing gate.

ux09Editing.browser.mjs brings critical paths from the standalone ux04Editing.browser.mjs into thirteen independent required cases. It reuses the existing native IndexedDB reader, Library/backup helpers, keyboard traversal and contrast calculation. No additional runner, dependency, retry, test timeout or suite budget is introduced.

Cases Production assertions
Six object variants Rectangle/ellipse room, polygon, river, note, token and stamp placement; actual keyboard primary-property edit, Apply and Cancel; complete atomic Undo/Redo; no-op Apply; deletion/Undo without stale selection; reload retains content and IDs
Object-specific guards Token picker cancellation creates nothing; note movement maintains tile references; polygon material edits preserve vertices; exact fractional river edits and point removal/Undo; locked stamps require committed unlocking before transforms
Overlapping regions Tiles, notes and stamps move together with retained IDs; unrelated rooms/rivers remain exact; fill/erase/Undo cover the whole selection; edge move controls are disabled
Gestures Native mouse stroke and one-step history; dispatched pointercancel/blur and real Escape discard unsaved previews; visible polygon completion; first token drag does not open a layout-shifting inspector before release
Scope and viewport Exact pan/zoom on Library return and same-tab reload; dirty forms cannot cross levels with equal object IDs, DM view or Library return
Preferences Stamp search/favorites survive reload; an injected localStorage write fault is announced, affects only current-visit favorites and does not block project editing or saving
Invalid fields and save failure Native required/range validation retains the original; quota failure rolls back every storage record; a real backup keeps applied unsaved object content and references; retry, Undo/Redo and reload preserve the intended project
Two keyboard layout cases Note chooser, multiline inspector edits, Apply, Cancel and dirty/clean Escape through real key navigation at 100% and 200% interface text; focus, clipping, hit testing, target size and contrast at seven viewport sizes

The layout matrix is 1440x900, 1024x768, 768x1024, 390x844, 844x390, 390x460 and 720x450 at DPR 1. The last two are reduced-height and reflow equivalents, not an OS keyboard or native browser zoom. Each engine records 14 layout observations. Apply has at least a 44x44 CSS-pixel target, an unobscured solid focus ring of at least 2 pixels, no document horizontal overflow, and no clipping ancestor or covering hit target. Measured focus width was 3 pixels; minimum sampled inspector label/help/action text contrast was 8.39:1 against the 4.5:1 assertion. These are scoped critical-control measurements, not full interface or WCAG conformance acceptance.

Defects found and fixed

The legacy inspector had a nested shrinking scroll area inside the contextual panel. At 844x390 with 200% interface text, keyboard focus reached Apply but the action was clipped and covered. The inspector now uses the contextual panel's single scroll area, as notes and initiative already do. In short viewports its action row is non-sticky, so the focused action can scroll fully into view. Before/after short-window and reduced-height screenshots were inspected; the production focus/hit-test assertion reproduces the original failure without weakening its threshold.

Firefox 153's native input.valueAsNumber converts the exact input string 0.3333333333333333 to 0.333333333333333. An isolated native input reproduced the difference. The shared inspector NumberField now parses the original number string, retaining JavaScript numeric precision while keeping blank required fields invalid rather than coercing them to zero. Two focused component regressions cover native-rounding independence and blank-field validation/cancellation. The same exact fractional browser assertion passes on all three engines; it was not replaced with a tolerance.

Revision-linked local evidence

Production assets at /Dungeon-Mapper/, macOS 26.6.2, locked headless engines:

Engine Browser version New editing cases Related existing cases Duration of selected runner
Chromium 151.0.7922.34 13 passed 4 passed 184.3 seconds
Firefox 153.0 13 passed 4 passed 180.3 seconds
WebKit 26.5 13 passed 4 passed 213.2 seconds

The 51 passing case/engine combinations include the existing shell, keyboard-critical, publication/preview and two-window session journeys. The unchanged UX-08 export/offline/controlled-worker journeys also passed all three engines, retaining player-export sentinels and real downloads. The production build, zero-warning lint and 41 focused editing/gesture component cases passed.

Evidence is retained in session 20233d90-7a79-4423-a564-f75af5b08662, files/editing-final-chromium, files/editing-final-firefox, files/editing-final-webkit and files/editing-final-exports. The three runner reports and journey attachments record the implementation revision above, browser version, results, per-layout measurements and page errors. Their preview servers used separate strict ports; each runner retained one worker and no retries. Earlier editing-keyboard-4 and editing-firefox artifacts retain the pre-fix product failures. Other earlier runs include harness corrections for native select typeahead, textarea accessible-name lookup, responsive-mode timing and absent-versus-undefined optional keys. Content equality uses the portable JSON contract; no-write assertions still compare raw native records and revisions.

This closes the bounded A-EDIT production gap, not every legacy scenario, all input devices or full accessibility qualification. Chromium pen/two-touch dispatch, physical touch/pen, assistive technologies, native zoom, forced-color qualification beyond existing styling, and participant acceptance are not promoted to passed by these cases. No project schema, renderer, audience policy or display/export contract changes. Continue with A-UPDATE's real application-version upgrade/recovery milestone.

September 28 local-only correction and reference retention

The September 16 result above is historical macOS evidence, not a passed Linux qualification. PR #185 is still open at dbba27f. In run 35037146541, Firefox passed 58 registered cases and failed the 200% inspector case at 390x460: Apply was partly outside the sheet and viewport. Build/test, Chromium and WebKit passed on that revision. The Browser qualification aggregate correctly failed because Firefox failed. The failure is an application focus defect exposed by the new A-EDIT coverage, not a package registry, runner or cancellation failure. Rerunning unchanged code is not the remedy.

The prior investigation left an uncommitted 390x440 regression size and a viewport-specific assertion message. Both are preserved. The original production assets reproduce clipping at that size locally in Firefox. The focus trap advances with focus(), which can leave a partly visible control clipped in Firefox. Keyboard traversal now also calls scrollIntoView({ block: 'nearest', inline: 'nearest' }), matching Notes and Initiative's existing focus-restoration pattern. Context panels reserve four pixels of scroll padding for the focus outline. No scripted scrolling is added to the browser test to conceal the defect, and no assertion, timeout or qualification requirement is weakened.

The correction is locally committed as aa40a14932537ace59c05536a5bd4fd5d99ae41b. A subsequent test-only commit, 0d0e0bb5fc1a73b5108da5ff25ae839d4f1d7112, includes both width and height in each draft and requires Apply to be enabled. This prevents adjacent 390x440 and 390x460 cases from accidentally exercising a no-op Apply. The retained eight-size matrix now has sixteen observations per engine.

Local evidence Result and applicability
Original application plus retained 390x440 case Firefox reproduced sheet clipping at 200%; command failed as expected, not passing evidence
aa40a14, macOS 26.7, DPR 1, production assets Four selected journeys per engine passed: both inspector text sizes, existing editor navigation and existing critical keyboard workflows; zero skipped cases
0d0e0bb, same application build Both strengthened inspector journeys passed on Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5; each size now commits a distinct edit; zero skipped cases
Focus trap, inspector and dialog component selection 31 tests passed, including forward/backward scroll-target assertions; jsdom is not visual scrolling evidence
Build and lint Production build and zero-warning lint passed; no dependencies were installed or changed

The final inspector results are in files/zero-actions-applied-{chromium,firefox,webkit}/. The September 28 host version above corrects an initially copied 26.6.2 label to the retained sw_vers result, 26.7. September 16's actual 26.6.2 records remain unchanged. Existing shell/critical-keyboard results are in files/zero-actions-final-{chromium,firefox,webkit}/. Each browser attachment records its source commit and browser version. All local runs kept the existing single-worker, no-retry and timeout settings. The inspector assertions still require unclipped, hit-testable, viewport-visible focus and at least 44x44 CSS-pixel actions. Minimum sampled contrast remained 8.39:1 and the focus outline 3 pixels. Selected reduced-height screenshots were inspected. No OS screen reader, native browser settings or physical-input tests ran.

The shared focus trap affects modal keyboard scrolling, so the existing critical-keyboard journey was refreshed alongside the inspector matrix. It includes nested dialogs, export-backup/recovery controls and session actions. Storage transactions, project schema, numeric parsing, renderer geometry, player projections and export pixels did not change in this correction. Their earlier evidence remains applicable to those unchanged implementations, but is not a new exact-head or release receipt. Renderer, preservation and UX-08 campaigns were not repeated solely because this CSS/hook changed. Future full hosted qualification is still required by the existing policy.

Build identity for both final local groups: App-CCpLUrr1.js, App-THvbwtv0.css, service worker SHA-256 0a1c061ddce052e4c2668a0541d8ac82feb680c7d8a06ee8ad4edb3b51fe5a9f. The original Linux report records the PR test-merge revision bd69b5f5cd70d3b7f66a6f25667b21c9d64d0ab5, based on main 72f32e1 and PR head dbba27f; it must not be confused with the local correction.

Reference dispositions, recorded before relocating the old untracked folders:

Material Disposition
Original Linux Firefox report, failed-case screenshot/trace, other case artifacts and UX-08 results in .pr185-check-fix-ci/ Retain the complete downloaded directory, unchanged, in session 20233d90-7a79-4423-a564-f75af5b08662, files/zero-actions-retained-ci/; do not commit generated reports
Earlier font/height/padding screenshots in .pr185-check-fix-probe/ Retain unchanged in files/zero-actions-retained-probes/ as diagnostic references only; screenshots without run/source metadata are not passing qualification
September 16 files/editing-final-* reports and synthetic fixtures in src/test/ Already retained; do not rerun unrelated storage, renderer or export campaigns or delete their evidence
New local reproduction Retain files/zero-actions-editing-baseline/, identified as dbba27f application code plus the retained 390x440 regression and diagnostics; this is a failed result
First corrected local run Retain files/zero-actions-editing-candidate/ as a working-tree result, not a hosted receipt

No original failure or synthetic fixture is discarded. The repository's existing policy remains: generated local evidence lives in session artifacts; CI evidence has fourteen-day retention; release evidence must outlive that window. Relocation preserves content, rather than replacing failure evidence with a successful report.

The relocation completed without removing evidence. The original Firefox browser-report.json SHA-256 before and after relocation is 80f03d45bc92c7f2fa7b17936f23bcb5ff07532ef6abfa5ab42cfaf2b4827b39. Read-only GitHub run metadata was added beside it, preserving both run-head and test-merge identities. The screenshots-only probes remain explicitly non-qualifying. No new evidence framework or test fixture was introduced.

Zero-Actions roadmap disposition and resume checklist

Budget and scope: September 28 task approval is 0 hosted runner minutes. This task initiated 0 hosted triggers, reserved 0 minutes and has 0 allowance remaining. No push, PR mutation, merge, workflow dispatch/rerun, hosted review/agent, paid-provider call or automation was requested. Local compute and disk usage were not metered; GitHub account quota, billed storage/cache usage and other actors' consumption are unknown. Read-only GitHub queries do not establish a free account balance. Previous run costs are historical, not spending authorized by this task.

Remaining work Owner / dependency / permission Reusable evidence and smallest next execution Completion evidence / stop
Locally completed: A-DOCS Agent; current UI and documented contracts; publication still requires budget d9036c5 updates README, Features, Sharing, help/demo guidance and release notes; targeted help/action cases passed See the milestone below. Do not repeat the audit or claim release/support approval
Bounded local pass: A-UPDATE Agent; owner still decides supported-version scope and final candidate; publication/hosted integration requires budget and policy scope Rebuilt deployment source 72f32e1, candidate d9036c5, manual harness db024ac; three-engine upgrade/rollback evidence below Reuse this pair's evidence where applicable. Rerun only for a changed build pair, schema or claim; final release evidence must name the accepted versions. Stop on lossy writes or an unresolved support decision
Owner: A-PRINT; digital A-ART decisions complete Owner/reviewer with approved art sheets, printer and print application; agree scale/alignment tolerances before measurement Digital art approvals closed September 29. Reuse current approved F07/A4/Letter exports and print actual-size overview/assembled miniature pages Measured paper scale/overlap and monochrome acceptance remain pre-release requirements. Do not repeat completed digital-art decisions or treat them as paper evidence
Local analysis plus owner: A-PERF Agent can profile; owner must identify/approve reference hardware and any support-scope exception Retain existing F05/cache diagnostics. Profile an identified workload only for an actual gap or proposed change; do not repeat the completed cache campaigns Measured target results or explicit scoped early-access decision. Development Mac results alone do not pass reference acceptance
Hosted: A-EDIT landing / #185 Agent plus owner numeric budget approval; final head and base must be refreshed; fix is only local Original failed run and current local correction/results above. After approval, one batched push and full existing CI on that head, including all three engines and UX-08; do not rerun the stale failed head or only the aggregate Build and test plus Browser qualification succeed with all constituent results on the actual candidate/merge revision. Stop on failure, quota rejection or budget exhaustion; diagnose before any retry
Hosted and remaining owner decisions: A-RELEASE Core local topology/artifact/cancellation implementation approved; durable retention and any additional human deployment approval remain open. Publication/post-merge work needs a numeric budget Local gating implementation and verifier/integration evidence below. Refresh the actual remote base, triggers and receipts before one batched publication; no deploy-only retries Exact deployed source matches successful qualification; versioned release evidence survives CI expiry; final-version rollback/recovery applicability resolved. Merge alone is not release approval
Later full-release: B-DEVICE / C-USERS Owner, owned devices and real participants; separate authorization/restoration plan for native settings and assistive tools Existing task scripts and synthetic fixtures, not AI personas or private data Actual platform/input/screen-reader observations and participant results. Explicitly deferred from early access, not passed or optional for full qualification
Optional: ART-05/09 and UX-10 Owner product and operating-cost decision Existing scope boundaries; no work started No implementation until explicitly selected; not initial-release dependencies

The subsequent zero-Actions A-DOCS/A-UPDATE milestones are recorded below. Remaining work needs the reference-hardware/workload decision for A-PERF, physical-print participation, or remaining A-RELEASE retention/approval decisions; hosted qualification still needs a numeric budget. Do not invent a supported browser/device range or imply that early access is approved. No dependency update is bundled: this failure involves focus/layout, not package exposure. Any future dependency finding must identify the installed path, affected version, runtime/build exposure and compatible fix before changing the lockfile.

Baseline CI waste assessment, before local implementation: Feature-branch pushes have no direct CI push trigger, but updating the open PR starts one full PR workflow. CI also runs after merge on main; Pages starts separately. There is no merge-queue trigger in the checked-in workflows. CI uses per-ref superseded-run cancellation; Pages has its own cancellation group. Each PR run performs four dependency installs/builds, one in Build and test and one in each engine job. Main adds a separate Pages install/build. Each engine already reuses its build for the runner and UX-08, so no extra build should be added between those steps. The WebKit pixel preflight deliberately runs before the full suite. Job limits remain 15 minutes for Build and test, 20 per engine and 2 for the aggregate; Pages has no explicit job timeout.

No change-aware selection currently exists. Deduplicating production builds would require artifact identity/integrity handling; safe docs-only selection would require an approved policy that keeps required checks fail-closed and distinguishes unexecuted evidence from passes. Adding paths-ignore, omitting an engine, making the aggregate green on a skipped job or removing the preflight is not an authorized optimization. Pages currently runs independently of successful CI, which remains an A-RELEASE defect to address under explicit policy scope, not a reason to weaken checks.

Historical run 35037146541 used about 32.5 aggregate job minutes, approximately 35 after per-job whole-minute rounding, not its 11.5-minute elapsed duration. Earlier main run 35034579302 used about 24.4 aggregate minutes, before this added editing suite; Pages run 35034579361 used 44 seconds across two jobs, approximately 2 rounded minutes. These are observed standard-Linux durations, not invoices. A future one-update/one-merge cycle with the enlarged suite is roughly 70-80 rounded runner minutes for two CI runs plus Pages, before retries; another full CI attempt adds roughly 35-40. Refresh durations, triggers, billing rules and artifact/storage implications, propose a conservative numeric ceiling with post-merge/retry reservations, and obtain approval before a trigger. None of these estimates grants an allowance now. Stop rather than exceed the approved ceiling.

Local documentation and real-version update milestone

September 28, 2026. All work is locally committed with no pushes or hosted triggers. Documentation/help source: d9036c503c833ca44061bef86283c13440e4d6b4. Manual rehearsal: db024ac7470e4f61dc7dcf6c557dcccf8e197a13.

A-DOCS reconciles README, Features, Sharing/demo scripts, the documentation index, changelog and in-app shortcut help against the actual UI source and existing UX-02/05/06/08 contracts. It corrects the obsolete fresh-Library G/Present flow, Rail/Tabs and mobile menus, import replacement claims, print shortcut, export-overlay labels, save promises and qualification overclaims. Library import previews and direct editor imports are described separately. Project backup JSON excludes separate session records; session recovery files have no dedicated importer. Historical GIFs remain labeled references. Nineteen targeted help/action component assertions, build and lint passed. Only help/action wording changed in application code; action handlers, publication and serialized data did not.

The prior source was identified read-only from successful Pages run 35034579361: 72f32e16c30be85e92f48e13fccb1576f36ed35e. It was exported with git archive into task-owned session artifacts, not another session or the main checkout. Both versions were built locally with Node 24.16.0 and the identical dependency lockfile (SHA-256 e038ea92ee0771faac9355d3892b01fc838daf42350251bbc7cbbde08806ee50). The archived build reused the existing dependency directory after byte equality of the lockfiles was checked. No install, upgrade or schema change was needed. This is a reconstruction of the deployed source, not a claim that the original hosted artifact was downloaded or its byte identity independently established.

Build App asset App SHA-256 Worker SHA-256
Prior 72f32e1 App-CbydfPzw.js 32c23606ba87c1b2ccd50ec93d1826f8d28dd179ba5f5d56e7b01d157c5a4af2 5cd81b01a4780d216c7126760b1f00255f17416780145782d1ea89f72502bd2c
Candidate d9036c5 App-BE003znV.js 171eaac1c1c34284f77cbc985b77e74a5a0190bc4554cdb8fa18606194bc8774 d1012ad893e5348c4b7e50b32809890492a8bc987c52f2fe4cc1d9020f4c8460

ux09Upgrade.spec.mjs uses the locked Playwright runner and real Vite production previews at the same /Dungeon-Mapper/ loopback origin. It restarts that origin with each complete build; it does not append worker comments or invent a data migration. Assertions verify distinct bundle bytes, the loaded document module, the actual App resource and content hashes after upgrade and rollback.

The synthetic input reuses ux09Library.browser.mjs's rich two-level fixture, including embedded image, custom theme/tile/stamp/template libraries and unknown extension fields, with one added player token and initiative entry. Public UI creates the local identity, recovery copy, session movement and named session checkpoint. There are no real users, live AI personas, external providers or native browser-setting operations.

Each engine asserts:

  • A held native IndexedDB write blocks the explicit update without navigation; the latest committed project then survives activation.
  • Every durable record remains equal immediately across upgrade and rollback, including project/session identities, revisions, source checkpoint, project recovery copy and named session checkpoint.
  • Private backup downloads before/after upgrade are equal; candidate session checkpoint restoration preserves the pre-restore progress in the recovery slot.
  • Candidate edits survive rollback; the older application can save a new project edit and session turn without losing assets, unknown fields or recovery history.
  • A real downloaded project backup imports as a separate identity, leaving source project/session work intact. The rolled-back application reloads and reopens that copy with the sole HTTP origin stopped.
Local environment Actual result
macOS 26.7, Chromium 151.0.7922.34 One complete upgrade/rollback journey passed, 19.9 seconds
macOS 26.7, Firefox 153.0 One complete upgrade/rollback journey passed, 22.1 seconds
macOS 26.7, WebKit 26.5 One complete upgrade/rollback journey passed, 19.6 seconds

All three passed in 62.4 seconds with zero skipped/flaky cases, one worker, no retries and the existing three-minute case/twenty-minute suite limits. Persistent profiles are synthetic and task-owned, and contexts/servers were closed by fixture teardown. Browser errors were empty. Neither browser-process restart nor OS interruption, physical storage loss, all historical version pairs, arbitrary downgrades or a schema migration was qualified.

Reference dispositions in session 20233d90-7a79-4423-a564-f75af5b08662:

Material Disposition
files/upgrade-prior-source/ Retained source archive, build and read-only deployment metadata; its dependency symlink is build tooling, not part of the immutable distribution
files/upgrade-candidate-dist/ Retained copy of the tested candidate distribution, checked against the full file-hash manifest
files/upgrade-final/ Retained runner reports, build/runner source identities, complete distribution hashes, synthetic persistent profiles, original input, native record snapshots and real project/session recovery downloads
files/upgrade-first-chromium/, upgrade-second-chromium/, upgrade-third-chromium/ Retained failed harness attempts: inherited preview server collision, then native Fetch response API mismatch. These failed before qualifying upgrade behavior and are not product failures or passes
files/upgrade-fourth-chromium/ Retained initial successful working-tree rehearsal; final three-engine evidence supersedes it for this pair
Existing UX-08, preservation and A-EDIT references Already retained. Their original outcomes remain unchanged; no unrelated renderer/export campaign was repeated

The manual config is deliberately outside default CI. The required runner still discovers 177 tests in nine files; this was a discovery check only, not 177 passes. No workflow, trigger, required-check policy, timeout or hosted resource changed. Commands and prerequisites are in Development.

Release applicability: This supplies bounded local A-UPDATE evidence for 72f32e1 -> d9036c5 -> 72f32e1, not an approved supported-version range or a final-release receipt. A-RELEASE must still bind the accepted candidate and supported predecessor to exact qualification and a publication/rollback policy. Adding this rehearsal to hosted CI or changing Pages qualification needs explicit policy scope and a numeric budget. The known #185 hosted Firefox failure is not cleared by these local runs. A-PERF reference acceptance, owner art approval, physical A4/Letter evidence and the later B/C gates remain open.

Accounting: The continuation retains a zero-minute hosted budget, zero initiated triggers and zero reservations. No remote publication or paid-provider call occurred. Local compute and disk costs were not metered; account-wide quota, billed storage/cache usage and other actors' usage remain unknown.

Local exact-artifact deployment implementation

September 28, 2026. The owner explicitly selected Approve local implementation for the design while preserving zero hosted minutes, no pushes and no remote-setting changes. Implementation source: 96b1b70ca09ca4610f7639d376c1b4f8cf3f9a80. A subsequent local test-only assertion also verifies dirty-checkout rejection; it does not change the production helper or the artifact-consumer evidence below.

The checked-in CI graph now builds the production distribution once after lint/unit success. Its canonical artifact binds the checkout, repository, event/ref, run/attempt, lockfile, tool versions and all distribution file hashes. Every browser engine downloads that exact immutable ID, checks its manifest against the producer's output hash, materializes a new dist, runs the unchanged qualification commands and checks the consumed files again. Renderer diagnostic bundles still use their existing in-memory compilation.

Build and test and Browser qualification keep their required names. The latter still fails on any non-successful required engine. Matrix fail-fast, lint ceiling, test selection, preflight, existing qualification timeouts and pixel/latency policies are unchanged. UX-08 still runs after another browser assertion fails, but not when no verified distribution was materialized. Playwright report metadata, copied manifests and UX-08 results retain the build identity. This adds no new runtime dependency and does not rebuild between suites.

Only a fully successful main push or main dispatch in this repository reaches Pages packaging/deployment. The separate deploy.yml is removed locally, including its manual bypass. Packaging and deployment verify the canonical bundle again. The deployment slot rechecks current main and the Pages artifact's ID/name/run/source/repository/expiry. Forks, PRs, off-main refs, stale sources and mixed attempts cannot pass those checks. Partial reruns require a complete budgeted rerun instead of reusing earlier-attempt qualification.

Qualification jobs cancel superseded work individually; there is no outer workflow cancellation to terminate the serialized Pages job. Pages writes and OIDC permissions are confined to that job. The existing main-only environment and server-side required checks were not modified. External cancellation or uncertain Pages completion still requires an operational hold, not an assumption that a remote request was undone.

Pre-submission intent and final receipt are distinct. A skipped submission records not-published; an unsuccessful Pages action records unknown server-side outcome; a successful action with a failed public-file check records published-but-unverified. HTTP checks compare qualified bytes with at most three rounds, a two-minute request deadline, ten-second request timeouts and bounded response sizes. There is no automatic redeployment. Canonical/qualification/ receipt evidence retains fourteen days; the intermediate Pages archive retains the prior one-day lifetime. Durable release retention remains unapproved.

Local evidence and applicability

Evidence Actual outcome and limit
Verifier/admission tests 34 tests passed, including real CLI exit codes, source/run/attempt/lockfile mismatches, mutated/missing/extra files, unsafe paths, links, stale main, fork/PR/tag/off-main denial, artifact metadata rejection, bounded smoke failures, receipt outcomes and dirty checkout rejection
Workflow validation actionlint 1.7.12 accepted ci.yml; existing lint passed. This is syntax/expression validation, not GitHub enforcement evidence
Build TypeScript/Vite production build passed. Application output remained unchanged by this tooling/pipeline slice
Canonical handoff integration Actual helper CLI created a bundle from the production build, materialized it back into the checkout and successfully verified all 18 files after selected existing production journeys
Selected WebKit journeys Editor navigation/focus/reflow and the floor direct-pixel case passed, zero skips; unchanged UX-08 export/offline/controlled-update journey passed on the materialized build
Hosted execution Not run. Artifact service behavior, Linux/Node 20 execution, job/concurrency semantics, environment permissions, Pages submission and public smoke remain unverified

Integration ran on macOS 26.7, Node 24.16.0, npm 11.13.0 and WebKit 26.5. Its canonical manifest SHA-256 is de4f799784c1c3cbdf7d187c0532ef920a05a7c6a4c0b9d555b445161d63e346. The source is the implementation commit above, but run/attempt/artifact IDs were deliberately synthetic (1) on the local feature ref. They are not a GitHub dispatch, artifact or hosted receipt. No Pages or GitHub publication API was called by the local tests; HTTP smoke tests used synthetic responses.

Retained references in session 20233d90-7a79-4423-a564-f75af5b08662:

  • files/gating-unit-final.json: final actual unit results.
  • files/gating-integration/: input distribution, canonical manifest/bundle, synthetic step outputs and production browser/export reports with matching manifest identity. Included for reproduction, not promoted to hosted evidence.
  • files/gating-actionlint-1.7.12/: pinned task-local validator and publisher checksums. It was downloaded only after the chosen actionlint command failed because the tool was missing; its archive checksum was verified. No global install or repository dependency change was made.
  • Existing A-EDIT, A-UPDATE, renderer, preservation and original Linux failure references remain retained. The application build did not change; their unchanged behavior evidence is not invalidated or relabeled as new CI success.

Local checks corrected a disallowed runner context in job-level env and the download action's artifact-ID extraction layout before publication. The action interfaces were inspected read-only at the exact selected versions. Verifier development failures were diagnosed without weakening assertions. No unrelated full browser/renderer campaign was repeated.

Hosted resume and stopping conditions

Core implementation approval is already recorded; do not ask for that same approval again. Before any push/dispatch/merge, refresh main, PR #185 and in-flight runs, inspect the final diff, estimate all jobs/attempts/post-merge work and obtain a numeric shared task budget. There are still zero approved minutes. The historical Firefox failure is not cleared by these local results.

The smallest normal hosted path is one batched PR update running full CI, then an independently authorized merge/main run with gating, packaging, deployment and public receipt. Retain the existing fail-closed checks and stop on any failure, skipped requirement, quota rejection, identity mismatch or uncertain deployment. Diagnose and request remaining/full-rerun budget before another trigger. Refresh old pending deployments before rollout; do not cancel unrelated runs or leave the old independent deploy path as a fallback. Historical reruns retain their original SHA/ref and privileges, so deleting deploy.yml does not prove that all legacy publication definitions are revoked. Cutover needs explicit owner authorization for any remote legacy-workflow restriction and verification of its actual limitations. This remains a release blocker, not a locally enforced repository-wide guarantee.

Do not automatically add the manual real-version rehearsal, new optional checks, or human environment protection to this scope. Final A-UPDATE applicability, durable release storage, any additional human deployment approval, A-PERF, owner art and physical printing remain separate pre-release decisions/evidence. Technical preview publication would still not authorize an early-access release.

Accounting: zero hosted triggers and zero reservations against the zero-minute budget. No pushes, PR mutations, merges, dispatches, reruns or remote settings changes occurred. Local compute, disk and validator-download transfer were not metered; account-wide quota, billed storage/cache and other actors' usage remain unknown.

September 29 local render-state investigation

Outcome: no renderer change retained. The owner authorized a bounded, zero-hosted performance investigation on the available Mac, not representative device acceptance or a performance-scope exception. Baseline: 2866790ba7e3b5904e8fa5c64e9b4e5c4868f6f0. Existing floor-cache and edge-cache results were reviewed first; this pass did not revive the rejected bitmap work.

The unchanged F05 v1 fixture and production diagnostic ran at 1440x900, DPR 1, unthrottled, on an Apple M5 Max / 64 GiB, macOS 26.7 (kernel 25.6.0), Node 24.16.0. Engines were Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5. Three repetitions per engine were run sequentially for the baseline and the initial prototype, avoiding concurrent test load. The host was not a reserved reference lab. Chromium repetition one includes CPU/timeline profiling; repetitions two and three do not. Other engines were unprofiled.

The Chromium baseline profile attributes approximately 3,702 ms inclusive to the main render effect over the recorded interaction sequence. Folio cached tile dispatch, including its direct non-floor fallback, accounts for about 1,949 ms; edge blending about 1,044 ms; lighting about 306 ms. These are sampled inclusive call-stack costs, not additive GPU timings or measurements of an isolated individual operation. Native Canvas execution and deferred work can affect attribution. Floor/edge raster drawing remains the main direction for investigation, rather than another speculative React/state cleanup.

The prototype grouped consecutive cached floor draws under one Canvas save/restore and absolute transform setup, flushing before other tile themes. It retained primitive order, the twelve-entry/56-path ceiling and existing edge-cache allocation bounds, with no new bitmap or map-state cache.

Engine / repetitions Paint p95 baseline / initial prototype (ms) Token p95 baseline / initial prototype (ms) Warm-ready baseline / initial prototype (s)
Chromium, unprofiled 2-3 144.3-149.0 / 146.0-147.5 144.3-144.4 / 141.9-145.8 1.50-1.99 / 1.49-1.52
Firefox, 1-3 148-149 / 142-143 141-165 / 137-143 1.23-1.26 / 1.24-1.26
WebKit, 1-3 76-86 / 77-79 83-94 / 76-84 1.46-1.50 / 1.51-1.51

These are ranges of individual repetition p95 values, not pooled percentiles. Chromium's profiled repetition changed paint/token p95 from 147.8/142.6 to 148.1/145.7 ms and warm-ready from 1.97 to 2.01 seconds. Chromium has no clear improvement. Firefox's paint measurements improved by 5-7 ms in this sample; WebKit token timings improved, while paint and warm-ready were mixed. No universal speedup, loading-target pass or representative-device result follows from these observations. F05 remains an event-to-render/frame proxy, not physical input-to-paint or INP.

The initial prototype passed the existing nine floor-render cases, including the 210-combination pixel matrix per engine and cache bounds. Expanded caller-state probes then found discrepancies under transformed, partially transparent multiply compositing. Chromium comparisons reached 2/255 RGB and 1-2/255 alpha difference. A later direct-path fallback for rotated/skewed/ mirrored callers removed those observed cases, but a WebKit fractional, nonuniform-scale case still reached 2/255 RGB against the unchanged 1/255 maximum. The required exact alpha and existing RGB/mean bounds were not relaxed. The additional cases are not passing evidence, and the precise native-raster contribution was not resolved within this bounded pass.

The modest and mixed timings did not justify more special cases or a fidelity exception. Later correctness refinements were not performance-qualified; do not attribute the initial prototype's timing numbers to the final retained rejection patch. The prototype and its API-specific tests were removed from the application together, not left behind with a weakened check. All six touched source/test files were restored exactly to baseline. A rebuild then matched all 18 production distribution files byte-for-byte against the retained baseline. The existing tests, artwork, renderer, memory ceilings and release targets are unchanged.

Retained in session 20233d90-7a79-4423-a564-f75af5b08662:

  • files/perf-sept29-baseline, perf-sept29-baseline-unprofiled and perf-sept29-baseline-other-engines: nine baseline repetition results, the Chromium profile/timeline, screenshots and environment/asset identities.
  • files/perf-sept29-candidate-profile, perf-sept29-candidate-unprofiled and perf-sept29-candidate-other-engines: nine initial-prototype repetition results. These are labeled dirty working-tree diagnostics, not committed candidates or release qualification.
  • files/perf-sept29-floor-candidate-pixels: original matrix passes; perf-sept29-floor-state-pixels: a harness export-placement error with six unexecuted cases, not a pass; perf-sept29-floor-state-pixels-fixed and perf-sept29-floor-fidelity-final: unresolved fidelity failures and traces.
  • files/perf-sept29-rejected-floor-batch.patch and perf-sept29-rejected-dist: final rejected implementation and build, retained for diagnosis rather than application. Patch SHA-256: 1c59d0fed51f2722f1c694f7a46efd4c61bdeb6b8aad99987b13b559cf1d94f9.
  • files/perf-sept29-comparison.json, perf-sept29-baseline-profile-summary.jsonl, perf-sept29-baseline-dist and perf-sept29-build-identities.json: per-run comparison, profile attribution, original files and exact restoration proof.

Fixture SHA-256: 35856d38cf71fdf0132c99f426d7552e0708234ccda2436276ac26d7fb3296b1. Unchanged diagnostic source SHA-256: 7e44eaa5122abf797d71f82cad36b3e01c979e7034d35c4eabc42f31b7b7a6f4. Earlier performance evidence remains historical, not overwritten. Retaining failed probes prevents rediscovering this rejected approach.

Next boundary: Reuse this profile to select a different measured approach, not another broad rerun. Avoiding full-map repaint during local token/gesture changes is a possible larger direction, but requires an explicit design for layer invalidation, derived geometry, fog, compositing and memory before implementation. It is outside this bounded pass. A-PRINT and reference/support decisions remain open; the 100 ms desktop, 150 ms mobile and two-second warm targets are unchanged and unaccepted.

Accounting: zero hosted triggers, zero reservations and zero approved runner minutes. No pushes, remote mutations, provider calls, dependency installs, subagents or native browser/OS setting changes. Local compute and disk usage were not metered; no account-wide allowance is assumed.

Token-drag partial repaint proposal

September 29, 2026, based on e1ba751 with the restored renderer. The owner asked to investigate avoiding full-map redraws after the rejected floor-state batching trial, then explicitly approved proceeding with the bounded prototype. After the first pixel mismatch was localized, the owner also approved expanding damage to complete overlapping artwork, while reiterating that no Actions minutes are available. The source trace itself needed no additional benchmark; the subsequent bounded implementation and local results are recorded below, not treated as release acceptance.

The owner subsequently approved conservative admission only where the native tile pitch is an integer number of device pixels. Other scales retain the full-quality full renderer, without a DPR whitelist, resolution reduction or tolerance change. Complete-overlap expansion also falls back at canvas edges, above 25% area, after sixteen closure passes, or beyond 2,048 temporary footprint records. These records are not retained raster surfaces.

Actual invalidation and drawing order

useCanvasEditingDraft stages token movement in a local map copy, changing the tokens array and the moved token's coordinates while sharing untouched branches. MapCanvas derives tiles from tiles/rooms/rivers, so that derivation can remain stable, but its main effect depends on the whole draft map and token arrays. Each movement therefore resets the entire backing canvas and redraws every layer. The minimap also depends on map, although it draws no tokens; that smaller cost is not the main target of this proposal.

The renderer draws background/paper, tiles, river banks, edge blending, grid, hand-drawn styling, atmosphere/lights, notes, structural overlays, annotations, markers, light icons, stamps, then tokens. Fog, FOV, stair badges and other interaction overlays can be drawn above tokens. A token-only canvas placed on top of the finished image would change occlusion or leave the original token behind. It is not a drop-in optimization.

During local drag previews, parent App still owns source-map FOV/exploration. Commit, cancellation and source-map changes have their existing gesture/history behavior. The optimization must not change when visibility, exploration, autosave or undo is updated. Do not move those calculations into the drawing fast path or suppress them for performance.

Recommendation and memory boundary

Prototype a same-canvas dirty-region repaint, not another full-map layer or a copied background patch. Clear and repaint the union of the last-painted and next token footprints on the original backing surface, preserving global coordinates and the normal layer order.

An extra F05-sized RGBA surface would add approximately 64 MiB at DPR 1, 256 MiB at DPR 2, or 576 MiB at DPR 3 before browser overhead. These are computed storage sizes, not measured process memory. A viewport-sized overlay costs less, but introduces coordinate/compositing and getCanvas() contract changes. Cached patch readback/blitting also risks synchronization and the cross-surface fidelity issues already observed. Neither is the first choice.

Proposed incremental retained raster budget: zero new canvases, ImageData caches or bitmap atlases. Keep the existing main surface, edge-cache ceiling and floor-path bounds. Retain only the last completed frame's input references, token identity/position and bounded rectangle metadata; release on scope changes/unmount. This is not a claim that browser command buffers, transient objects or GPU memory are constant. Include long-drag observations before making any memory claim.

First prototype scope

Condition Proposed disposition
Active move-token drag preview in trusted Edit, before commit Candidate for partial repaint
One token changes only X/Y; IDs, order and every other token field are unchanged Required; detect against the last successfully painted frame, not the last pointer event
Existing complete frame, unchanged project/level/source, dimensions, tile size, DPR, non-token map branches and drawing inputs Required; unknown/additional input changes default to full repaint
Fog enabled but not drawn because GM Show Fog is off Can remain eligible if all visibility inputs are unchanged, as in the existing F05 editor workload
Visible fog, player/DM-view mode, FOV inspection, print mode or competing gesture overlays Full repaint
Initial scope: Folio tiles/materials, bundled known token paths, existing paper/edge/lighting effects Supported candidate set; preserve every effect and rendering order
Imported background images, custom themes/stamp images, non-Folio tile overrides, hand-drawn effects, unknown or text/emoji moved-token glyphs Full repaint for the initial proof, not silently reduced detail
Commit, cancellation, lost capture, blur, tool/mode/level/project change, undo/redo, async asset readiness or resize Full repaint and reset partial-frame eligibility
Invalid bounds or a damage rectangle exceeding 25% of the canvas Full repaint; this is a proposed optimization cutoff, not a map-size/support limit

Exclude unsupported artwork before entering the fast path. The ordinary renderer continues handling it unchanged. Cache the eligibility result only while its immutable source inputs remain identical, rather than scanning every tile on every pointer movement. Keep a conservative comparison of the full external render inputs and relevant local drawing state; do not create a second incomplete dependency whitelist that can miss future changes.

Repainting contract

Damage includes both the old and new token footprints, stroke width and selection ring, not just the occupied cells. For the existing one-cell token at a 32-pixel tile size, the selected outer radius is 18.56 pixels, which extends past the 16-pixel half-cell. Unknown glyph extents must fall back, not guess. Round outward in physical pixels, add an antialiasing gutter, clamp to the canvas and merge into one rectangle to avoid painting alpha twice.

As a work-count illustration only, a selected one-cell token moving one cell horizontally in F05 gives a conservative 4x3-tile damage box with a two-physical- pixel gutter at DPR 1. A one-cell drawing halo would visit up to 30 tiles instead of 16,384. This is neither a latency prediction nor an allocated patch. The halo must be justified against the admitted primitives, not assumed safe for arbitrary imported art or stroke settings.

The initial full frame retains the current backing-size initialization. A partial frame must not assign canvas width/height, which would erase all unaffected pixels. It must begin from the same known paint state, clip using physical-pixel-aligned bounds on the original surface, clear that rectangle, then replay the ordinary scene paint body under the clip. Full and partial rendering must share the paint order rather than duplicate the renderer. If arranging that seam needs a broader refactor than this bounded proof, stop and revise the scope before implementing it.

Cull only loops whose footprints are established: tiles, local river-bank cells, edge bands and ambient-occlusion cells, using a conservative halo where needed. Neighbor/material queries still address the entire source map. Keep global artwork seeds and full-map dimensions for texture/gradient/cache configuration. Do not translate to a patch-local origin or call the export renderer to manufacture a replacement editor patch.

Replay the remaining scene layers in their original order under the clip, including unchanged overlapping tokens. A moved token must not jump ahead of stationary tokens, stamps, stair badges or selection overlays. Do not assume the area beneath it is plain floor. Preserve the complete getCanvas() result, hit testing, DOM/accessibility summaries and existing export behavior.

The edge cache's populate-before-sample behavior must remain intact. New range arguments, if needed, default to full-map traversal for existing callers. Print/export and the separate Run/player surfaces remain outside this first prototype. Do not add another canvas cache or browser-specific pixel exception to rescue a failed implementation.

Evidence required before retaining an implementation

Use the actual editor's full-frame paint path as the oracle, not the different export renderer. Compare complete equal-sized surfaces after each move, including unchanged pixels outside the damage, overlapping tokens, old-position erasure and boundary seams. Cover 1/2/3-cell tokens, selected rings, map edges, coalesced/long jumps, mixed materials, vectors, furnishings, lights, DPR 1/1.25/1.5/2/3 and fallback scenes. Exact alpha and the existing floor RGB/mean bounds remain the fidelity standard; no perceptual-only substitution.

Exercise start/commit/cancel, return-to-origin, undo, source replacement, fog/FOV/mode/print transitions and asset readiness. A rejected/unsupported fast path must perform the full render, not return a success-shaped no-op. Keep all pixel reads in qualification rather than production rendering.

Measurement dependency: F05 currently identifies a draw through backing- width assignment. Keeping that unchanged would miss successful partial paints and hang the diagnostic. Extend its test-only observation to main-canvas clearRect as well as width reset, coalescing the synchronous paint stack and retaining the two-frame completion rule. Observe real pixel-mutating work, not only a pointer event or context lookup. Extend the delayed-draw probe to cover partial clearing and reject unrelated-canvas/context-only false signals. Keep the fixture, sample counts, undo/persistence assertions, retries and timeouts.

Run a small fidelity proof first. Only if it passes, capture a matched baseline and candidate using the same measurement probe, browser, viewport, DPR and unthrottled host conditions. Compare profiled/unprofiled repetitions separately and retain all results, including paint and warm-ready regressions. The target is a clear repeatable token-drag improvement, not a timing assertion tuned to this Mac. Painting gains, physical input latency and representative-device acceptance are not implied.

Decision and stopping point

The owner approved this token-preview-only, zero-additional-raster prototype and the bounded overlap extension. This does not approve a general layer/cache redesign, visible-fog optimization, paint-stroke optimization, larger memory budget, raised tolerance or hosted run.

If the prototype cannot preserve the existing scene with these constraints, retain the failure and stop rather than accumulating special cases. A later layer/invalidation design requires a new scope decision. Current production code was unchanged during the initial design. Subsequent implementation remains local-only with zero hosted triggers/reservations, no agents and no native setting changes. The result section below records why the bounded implementation was retained; the wider layer/cache redesign is not approved.

Owner-approved release-candidate comparison policy

October 1, 2026. The owner wants the core application updated and available for real user testing as soon as practical, using that feedback to reach full release rather than waiting for every optional roadmap item. After receiving the specific tradeoff, the owner explicitly approved removing only the single-pixel RGB maximum requirement from the retired optimization's full-vs-full comparison.

This is an informed acceptance-policy change, not a source fix for the underlying native-raster discrepancy and not an infrastructure-rerun claim. The previously failed WebKit run remains failed under its original policy. This decision does not grant blanket permission to relax other tests.

Exact scope

In src/test/tokenRepaint.spec.mjs, the complete-frame comparison keeps:

  • exact alpha equality;
  • its existing whole-image mean difference limit of 0.01 byte levels per channel, or 0.01/255 normalized;
  • zero candidate partial draws and a newly observed complete draw at every comparison;
  • all six DPRs, tile sizes, token footprints, overlapping artwork and former fallback scenes;
  • movement, cancellation, commit-position/reference-state checks and the eighty-move no-new-canvas allocation exercise.

Only maximum <= 1 is no longer an acceptance condition in this test. Maximum/worst-pixel observations remain in the attached evidence. Focused policy controls exercise the retained mean boundary and reject alpha changes or mean overflow. The RGB maximum is diagnostic rather than silently omitted. Every other pixel/art, privacy/fog, editing, export, data-preservation and required CI check is unchanged, as are retries, timeouts and workflow topology.

Residual risk is explicit: isolated color differences can now pass this comparison when the whole image and transparency stay within its remaining limits. It no longer promises per-pixel RGB equality. This is considered proportionate for two ordinary full redraws after withdrawal of the optimization. Any future partial renderer still requires a separately approved validation plan and cannot bypass the full-draw assertions or inherit permission to reintroduce itself.

Candidate-first execution order

  1. Consolidate the current core-app fixes, permanent owner guidance and this approved policy update into one candidate. Do not add optional features.
  2. Obtain any necessary publication/CI budget scope, then require the complete agreed exact-head checks and existing merge gates. Verify the qualified Pages deployment before sharing an updated hosted candidate.
  3. Run a small first round with real DMs, using new test maps or independent copies and downloadable backups. Cover create/edit/undo, save/reload, backup/reimport, player-safe display and export. Capture blockers and the highest-value friction, not an elaborate new analytics system.
  4. Fix material issues and feed results into the remaining release decisions. Defer live GitHub hosting, collaboration, extra art and further micro- optimization unless user feedback establishes a real need.

This ordering makes user testing part of reaching release, not a reward for finishing the entire roadmap. It does not assert general availability, waive the existing physical-print gate for a release that promises printing, claim representative hardware/accessibility qualification, or approve a merge. The eventual supported release scope must match its actual evidence and any explicit owner-approved deferrals.

Local result and exact-source identities are retained in files/pr185-release-candidate-policy-source.json and files/pr185-release-candidate-policy/ in session 20233d90-7a79-4423-a564-f75af5b08662. A new hosted result must qualify the candidate under this explicitly revised policy; local results are not relabeled as Linux evidence. No additional hosted trigger or budget increase is implied by this test-policy approval.

Owner value judgment: proportionate quality work

October 1, 2026. The owner explicitly asked for a permanent project and cross-project preference: always recommend against this kind of work when the expected value is so small. Future agents should make that judgment proactively, not wait for the owner to interrupt a costly investigation. It applies to features, optimizations, testing, polish and maintenance, not only this rendering incident.

Before another diagnostic cycle, refactor or budget request, state the likely user benefit, material risk, uncertainty and incremental cost. If the remaining benefit is negligible, recommend stopping or deferring and suggest a simpler, proportionate alternative. An available or previously approved budget is not a reason to continue. A red automated check establishes a qualification blocker, but does not by itself establish that unlimited engineering effort is worthwhile or that its original acceptance criterion still fits the product.

Concrete example and limits of the evidence

After the owner-approved full-redraw restoration at 3cb439b, hosted run 36920945164 passed Build and test, Chromium and Firefox. Linux WebKit passed 62 cases and failed the specialized token-repaint comparison after 528 successful pixel observations. At DPR 3, one pixel out of 2,985,984 differed between two complete redraws: maximum RGB difference 23 versus the required limit of 1, alpha difference zero, and mean difference about 0.0000041862. Both render counters confirmed two full draws and zero partial draws.

Two bounded diagnostic repetitions on the development Mac captured 8,809 drawing commands/state observations per canvas with identical signatures and matching pixels. Instrumentation changes timing and does not prove what Linux did. The underlying draw-input, harness or native-raster cause remains unisolated. No material editing, map geometry, player-safety, data-preservation or export defect was demonstrated by this isolated observation. That is not proof of harmlessness, and the original failure must not be erased.

The owner does not want further expensive perfection chasing for such a small expected benefit. Keep the simpler full-redraw behavior. Recommend reviewing the specialized comparison's purpose now that the optimization it was created to validate is absent, instead of reflexively asking for another costly Linux run. Preserve meaningful rendering, fog, export, interaction and data-integrity coverage. Small differences are not automatically low-value when they affect those guarantees.

Retained evidence is in session 20233d90-7a79-4423-a564-f75af5b08662, under:

  • files/pr185-ci-36920945164-failed.log;
  • files/pr185-ci-36920945164-webkit-artifacts/;
  • files/pr185-ci-36920945164-pixel-disposition.json;
  • files/pr185-ci-36920945164-local-draw-commands.json.

The broader PR qualification task has used approximately 97 rounded standard Linux runner minutes across three runs, not all attributable to this pixel discrepancy. No further diagnostic run was started. The proposed extra run and increase to a 300-minute ceiling were not approved and should not be treated as the recommended default next step.

Authorization boundary

At this decision, this was a durable prioritization and recommendation policy, not approval to relax, replace or remove the current comparison, skip required CI, alter budgets, declare the failure fixed, or merge without qualification. A narrow test-policy change still needs an explicit, informed owner decision with its retained coverage and residual risk stated. Until then, the required check remained failed and publication remained held. The later specific approval above changes only the named RGB-maximum condition, not these general safeguards.

The cross-project preference is also saved in the owner's personal Copilot instructions. Repository-scoped guidance is in .github/copilot-instructions.md, and the roadmap links this decision so future work does not repeat the same low-value escalation.

Owner-approved full token redraw restoration

October 1, 2026. The owner explicitly selected Restore full redraws after reviewing the correctness/performance tradeoff, then requested complete documentation. This withdraws the e55bc1e regional token-preview path from production across all browsers, not only the failing platform. It is a local correction pending a separately authorized push and fresh hosted CI.

Why it was withdrawn

The WebKit job in run 36914937106 tested published source 96757fedaeb0afcf491216019e5d1ca61eba1a99 on ubuntu-latest. The required token-only repaint matches the complete editor across movement and cancellation case failed at move-5-5, DPR 1, 24-pixel tiles and a one-cell token with fog/text fallback disabled. The candidate had one full draw and one partial draw; its independent full-frame reference had two full draws.

Maximum RGB difference was 3 byte levels, exceeding the unchanged limit of 1. Alpha difference was zero and mean difference was about 0.002526, within the 0.01 limit; passing the mean does not excuse the failed maximum. The worst pixel at (80, 48) was [159,144,118,255] versus [160,147,120,255]. This was a PR-caused fidelity failure in the new partial path, not an established registry/runner flake or a defect in the fail-closed Browser qualification aggregate.

A focused macOS WebKit attempt at local 80b1b31 passed all 632 comparisons. That did not reproduce or disprove the Linux defect. The exact regional clipping/compositing cause was not isolated. The owner chose to remove the unqualified fast path instead of spending another diagnostic cycle trying to retain it. The original evidence remains unchanged in:

  • files/pr185-browser-36914937106-webkit-disposition-80b1b31.json;
  • files/pr185-browser-36914937106-webkit-failed-xYpx3v;
  • files/pr185-browser-36914937106-webkit-aggregate-sDXgRb;
  • files/pr185-browser-36914937106-webkit-repaint-local-cAO4TQ.

The two separate file-input ambiguities in that workflow were already fixed locally by 80b1b31, using the exact Import project label in the audience and session journeys. That selector correction and this renderer correction must both be included in the next candidate. No rerun of unchanged 96757fe would establish either fix.

Implementation and preserved behavior

MapCanvas now unconditionally resets the main canvas dimensions for each required frame, then runs its existing complete painter in the existing layer order. Last-painted-frame metadata, partial-scene eligibility, damage planning and regional clipping were removed from its production render path. Tile, river-bank, edge-blending and lighting traversal use the complete map, without region bounds. There is no platform detection, eligibility exception, hidden switch or resolution/detail reduction.

The canvas itself is reused; no additional retained bitmap/canvas cache was introduced. Coordinate-only hover suppression, ordinary input comparison, bounded edge strips and Folio floor-path caches are preserved. The pure damage/scene-bound helpers and their unit tests remain as experimental history but are not called by the renderer. sameRepaintInputs remains in use for its existing prop-change comparison. Drag state, token selection, move commit, cancellation, undo, persistence, player projection and export logic were not rewritten. The restoration changes the painting strategy, not map contents.

The required pixel test remains registered under its original name, with the same six DPRs, 20/24/32-pixel tiles, token footprints, overlapping artwork, edge movements, return-to-origin, cancel/commit and eighty-move allocation exercise. The visual assertions still require alpha equality, maximum RGB difference at most 1 and mean difference at most 0.01. No tolerance, case, timeout, retry or required check was removed or weakened.

The former partial >= 4 activation assertion is intentionally replaced: the owner-approved behavior no longer performs partial paints. Every comparison now requires the candidate's partial count to remain zero and its complete-draw count to increase since the previous observation. This proves the restored production path actually drew each sampled frame instead of silently accepting stale pixels or just suppressing the failing comparison. A component regression also verifies complete canvas resets for token edits in a formerly eligible Folio scene.

Performance tradeoff and reintroduction gate

The former eligible-preview gains, roughly 143-146 to 22-23 ms in Chromium and 141-143 to 22-23 ms in Firefox on the September 29 development Mac, are historical measurements of the withdrawn implementation. They are not current results. Full redraws can make token dragging less responsive on dense maps. The owner accepted that tradeoff to prefer correct map pixels. No fresh broad F05 performance campaign was run for this correction, and no exact current slowdown, representative-device pass or release-scope exception is claimed. A-PERF remains open.

Any future partial-redraw proposal must be separately scoped. Before it can return to production it must address the retained Linux WebKit reproducer, preserve these exact pixel and allocation requirements across the full matrix, pass required exact-head hosted qualification for all three engines, and show source/build-matched performance benefit without hiding failures behind an OS exception or widening tolerances. Reuse the retained failures and baseline instead of discarding them. New hosted experiments need their own approved trigger scope and accounting; the current CI budget is not an unlimited optimization allowance.

Local evidence and publication boundary

The corrected working tree was based on 80b1b31. Final source and built-file hashes are retained in files/pr185-full-redraw-source.json; browser reports marked as working-tree runs are not relabeled as hosted receipts.

Evidence Result
files/pr185-full-redraw-unit.json 37 selected component/planner/cache cases passed, including the new complete-reset regression
files/pr185-full-redraw-browser/ Three engines passed, 632 complete-frame comparisons each, 1,896 total; candidate partial draws zero throughout
Chromium pixel maxima RGB 0, alpha 0, mean 0
Firefox pixel maxima RGB 0, alpha 0, mean 0
WebKit pixel maxima RGB 1, alpha 0, mean 0.0000015070408950617283
files/pr185-full-redraw-production/ Six native production journeys passed: token placement/affiliation/player export and mouse cancellation/first-token drag on each engine
Lint / production build Passed; existing App chunk-size advisory remains visible

The local browsers are Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5 on the development Mac. Synthetic pixel comparisons and native production journeys are distinguished; neither is a fresh Linux receipt, physical-input latency measurement or release approval. All local checks used the existing single-worker, no-retry and timeout settings, with no skipped cases. Owned preview/browser processes were stopped by the runners.

No push or hosted job was started for this local restoration. Observed rounded usage across the two completed PR runs is 49 minutes of the approved 200, with 100 conditionally reserved for eventual main/Pages and 51 unallocated. The two approved PR-update triggers are consumed. A third push/run needs explicit additional scope and sufficient ceiling before publication; no merge, reviewer request or automatic retry is authorized. Actual billing, storage and account balance remain independently unverified.

Token-drag partial repaint implementation and local results

Historical implementation, withdrawn from production on October 1, 2026. The source, measurements and failed intermediate experiments below are preserved, not rewritten as passing Linux evidence. See the owner-approved restoration for current behavior and the performance tradeoff.

September 29, 2026. Originally retained locally at e55bc1e358780b7ba44110ffb0702eeccb5fca79, following the explicitly approved prototype, complete-overlap expansion and conservative native-grid rule. The subsequent test-only tile-size expansion and comment punctuation cleanup were rebuilt and verified against all eighteen retained distribution hashes: the measured application bytes are unchanged. No branch was pushed and no hosted receipt is claimed.

What changed

Trusted Edit token-drag previews can repaint a bounded area of the existing main canvas instead of resizing and repainting the whole map. The planner compares the last completed frame with current inputs, admits one token moving only X/Y, and expands the damage to complete intersecting artwork. Tile, river-bank, edge and ambient-occlusion loops receive conservative cell bounds; neighbor queries and all world coordinates remain global. Other scene layers, including overlapping tokens, keep their existing order under the clip. The common paint body is unchanged apart from those bounds and its outer save/restore scope; most of the large MapCanvas diff is indentation.

There is no new retained canvas, bitmap or ImageData cache. The existing edge/floor caches keep their bounds. Temporary footprint metadata is capped at 2,048 records and sixteen closure passes; a region above 25% of the canvas, touching its boundary, or failing eligibility uses a full redraw. Normal source updates, commit/cancel, unknown glyphs, custom/imported artwork, visible fog/FOV, DM/player view and print remain on the full path. No movement, visibility, history, autosave, token ordering, export or artwork semantics changed.

Admission uses the native Canvas transform's tile pitch, not a guessed DPR whitelist: tile size times native scale must be integral in device pixels. An unaligned grid uses the unchanged full-quality renderer. Fractional-scale cases remain in the pixel tests, with fallback asserted; there is no resolution reduction, high-DPR detail cutoff or widened pixel allowance.

Why the extra bounds and fallback were necessary

The first token-footprint-only implementation differed by up to 16/255 when a clip crossed existing artwork. Text omission and full traversal under the same small clip did not remove the mismatch; a full-surface clip control matched. Complete-object expansion resolved that case. Canvas-boundary and nonintegral device-grid cases still exposed 2/255 rounding differences. The final rule keeps those cases on full repaint rather than accepting a different picture. The owner approved this narrower optimization eligibility explicitly.

Those failed experiments are retained, not reclassified as passes. Early harness setup also used incorrect fixture dimensions and initially exercised no movement; the commit-position assertion caught that, and strict harness type-checking plus explicit partial/full counters now prevent a false proof. No timing was used to override a fidelity failure.

Matched local measurements

Baseline application 2866790ba7e3b5904e8fa5c64e9b4e5c4868f6f0 was served from its retained distribution, not rebuilt with candidate source. Candidate application is e55bc1e. Both used the same updated F05 probe blob, unchanged F05 v1 fixture, trusted input, sample counts, persistence/undo assertions, single worker, no retries and existing timeout budgets. Each engine completed three baseline and three candidate repetitions, run sequentially.

Environment: Apple M5 Max, 64 GiB, macOS 26.7, Node 24.16.0, 1440x900, DPR 1, no throttling. Chromium 151.0.7922.34, Firefox 153.0, WebKit 26.5. This is the available development Mac, not approved representative hardware.

Engine / repetitions Token-preview p95 before / after (ms) Paint p95 before / after (ms) Warm-ready before / after (s)
Chromium, unprofiled 2-3 143.0-145.7 / 21.5-23.2 146.4-149.0 / 148.9-150.1 1.51 / 1.50-2.02
Firefox, 1-3 141-143 / 22-23 142-144 / 142-143 1.24 / 1.22-1.26
WebKit, 1-3 80-94 / 63-79 81-86 / 80-87 1.47-1.61 / 1.52-1.59

These are ranges of each repetition's p95, not pooled statistics. Chromium's profiled repetition changed token p95 from 147.5 to 24.9 ms, paint from 154.7 to 151.5 ms, and warm-ready from 2.01 to 2.02 seconds. All twelve token-drag samples in every candidate repetition used partial painting; every baseline token-drag sample used full painting. The clear improvement is scoped to eligible editor previews, not the separate Run workspace or all dragging.

Release/commit still redraws the full map. Its single observation per repetition was 145.0-153.3 / 147.6-152.2 ms in Chromium, 141-144 / 142-148 in Firefox, and 81-82 / 82-87 in WebKit. Do not call these single observations tail-latency qualification. Paint and warm-ready results are mixed; there is no claimed paint/loading improvement. Small full-render regressions and the variable two-second Chromium readiness result are retained rather than averaged away.

The probe now observes actual main-canvas width resets or regional clears through the end of the synchronous drawing stack and two frame opportunities. Delayed full/partial probes reject early frame, context-only and unrelated- canvas false signals. This remains an event-to-render/frame proxy, not physical input latency, INP, or a pass of the unchanged 100 ms desktop/150 ms mobile/ two-second warm-launch release criteria.

Correctness and cost evidence

The in-memory browser harness mounts the actual MapCanvas twice. A changing, nonvisual callback makes the reference take its ordinary full-render fallback; the candidate uses partial painting. Complete equal-sized surfaces are compared, not cropped patches and not an export-renderer substitute. The final expanded matrix covers 20/24/32-pixel tiles, six DPR values (1, 1.25, 1.3, 1.5, 2, 3), 1/2/3-cell tokens, overlapping art, map edges, long moves, return-to-origin, cancellation, commit and visible-fog/text-glyph fallbacks. These comparisons use synthetic pointer events; native production journeys and F05 separately cover genuine input and persistence.

The alpha-equality and maximum 1/255 RGB / 0.01 mean bounds remain intact. The expanded matrix records 632 complete-frame comparisons per engine: maximum RGB difference 1/255, alpha difference zero, and maximum mean 0.000264/255 in Chromium/Firefox and 0.00000151/255 in WebKit. Each engine also completes an eighty-move drag/cancel case with no new DOM canvas elements observed. That does not measure total GPU/process memory or qualify long-running physical-device behavior.

Local evidence includes 117 targeted unit/component/audit tests, strict browser-harness type-checking, zero-warning lint, all three expanded pixel matrices and six delayed-draw probes. Existing native editing cancellation and viewport/scope journeys passed on all three engines. The unchanged F05 functional assertions passed in all eighteen matched repetitions.

Build/type-check succeeds, but the App chunk is now about 506.68 kB minified (148.95 kB gzip), versus 499.14 kB (146.37 kB gzip) in the baseline. Vite's 500 kB advisory is visible; its threshold was not raised or hidden. No unrelated code-splitting refactor or dependency upgrade was added.

The required runner discovers 183 tests in ten files after adding the editor-comparison case and second probe per engine. Discovery is not execution: the whole required hosted suite was not run. Required check names, engines, pixel tolerances, retry/timeout policy and fail-closed aggregation are unchanged. Future CI budgeting must account for the added comparison case; do not assume local rendering savings mean fewer total hosted minutes.

Retention and resume

Session 20233d90-7a79-4423-a564-f75af5b08662 retains:

  • files/token-region-matched-baseline and token-region-matched-candidate: all eighteen measurements, browser/host/source/probe identities and Chromium profiles. files/token-region-matched-comparison.json retains individual runs.
  • files/perf-sept29-baseline-dist, token-region-final-dist and token-region-build-identities.json: complete builds and file hashes. Baseline App hash: 18acacd9d1e40bc42de5ef21de819f5a80cec5a7542846e98e26fd15ac4a3310; candidate App hash: a3b8e8e21a2f1999b196acbcef08124a7d79238552d4b92b9b95060e4ad29852.
  • files/token-region-qualified-local, token-region-final-size-fidelity and token-region-unit-final.json: clean-revision proof/probes, expanded test-only tile-size proof and unit results. Earlier native gesture results are in token-region-final-fidelity, which also retains the fractional-grid failures and must not be represented as an entirely successful run.
  • files/token-region-first-pixels, token-region-third-pixels, token-region-localized-pixels, token-region-opaque-pixels, token-region-overlap-pixels and token-region-aligned-fidelity: rejected intermediate outcomes. token-repaint-*-diagnostic.json and the retained diagnostic script explain the text/stamp/culling/full-clip isolation; these altered-fixture controls are not qualification.
  • files/token-performance-baseline.config.mjs: the local-only configuration that serves the retained baseline with the common current measurement probe.

Keep the earlier rejected floor-batch evidence and all artwork approvals. No cleanup discarded an original failure or synthetic reference. The next hosted step remains one separately budgeted, batched publication and exact-head qualification. No performance-target exception or release approval was granted.

Accounting: zero hosted triggers/reservations against zero approved runner minutes. No pushes, PR mutations, workflow dispatches/reruns, deployments, paid-provider calls, native browser/OS setting changes or agents. Local compute and disk were not metered; account-wide usage remains unknown.

Dense F05 export and cancellation milestone

September 30, 2026. The owner authorized a bounded independent local export check, without physical printing or Actions minutes. Application baseline: 7b5dd9768fe090b8521d336cc83ea51697208462. The observed feedback fix is c7724e7d3f7b3eb5ded71cade3f5aa5b1fb36a39; final cross-browser harness: 8bea2e06c9d305fb24f3d25bf95776631338dcc9.

The new journey reuses unchanged F05 v1: 128x128 cells, 100 tokens, 200 furnishings, 100 notes, sixteen room shapes, a river, paper, edge effects and lighting. Existing UX-08 coverage for simpler page plans is retained rather than repeated. This case checks selected dense pages and interruption, not completion of a 252-page batch.

Reproduced issue and surgical fix

The initial Chromium run accepted cancellation promptly but took about 1,033 ms after receiving it to show the final cancelled message while a native PNG encode finished. The main thread remained responsive during most of that wait. ExportDialog now acknowledges the request immediately with Cancelling export. Waiting for the current operation to finish. The cancellation button becomes disabled and output/new-export controls remain locked until the current operation settles and its normal cleanup completes.

No native encoding is claimed to be preempted. The implementation does not race an abort against encoding and then allow overlapping export jobs. Completed downloads are kept, later requests are suppressed, and the original allocation limits, renderer, PNG metadata and project data contracts are unchanged.

A held-operation component regression verifies immediate acknowledgement, continued locking, final cancellation and a fresh usable controller on retry. Fifteen targeted component/export tests, build and zero-warning lint passed. The pre-existing App chunk-size advisory remains visible and unchanged in policy.

Actual local coverage and limits

The final case renders and encodes the first two color DM pages of the 300-DPI Letter plan. It retains the first real download and holds delivery of the second real encoding callback. Native Enter on the focused Cancel button must show pending cancellation and disabled controls before the callback is released. The second page must not request a download. This is an explicit injected pending-operation case, not natural cancellation-latency evidence.

The case then injects one null encoding result for page 252, verifies the visible error and cleanup, retries the actual last page, and downloads a player-facing 2048x2048 image at sixteen pixels per cell. Every completed file is retained, browser-decoded and checked for dimensions, PNG chunk boundaries, end marker, physical-resolution metadata and nonblank content. Actual private backups and all IndexedDB map/recovery records must remain equal before/after. There is no new claim about audience policy beyond the existing projection coverage; this test is not a substitute for pixel/sentinel qualification.

Controlled local result Chromium 151.0.7922.34 Firefox 153.0 WebKit 26.5
Outcome Passed Passed Passed
Largest observed export canvas 8,415,000 pixels 8,415,000 pixels 8,415,000 pixels
Peak combined managed page/content capacity 15,165,000 pixels 15,165,000 pixels 15,165,000 pixels
Managed surfaces explicitly zeroed after completion/cancel/error Yes Yes Yes
Temporary PNG URLs revoked All three All three All three
Completed PNGs decoded First page, last page, player image Same Same

The combined capacity is about 57.85 MiB at four bytes per pixel, not total browser memory. Observation covers HTML canvas dimension assignments and the known page/content surfaces, excluding the editor, pre-existing artwork caches, native encoding/GPU buffers and verifier-image decoding. It rejects an over-limit assignment before allocating it, retains the 16 MP / 8192-side per-surface bounds, and prevents a fourth batch download if the audit runs away. No limit, timeout or assertion was raised to obtain a passing result.

Natural post-fix Chromium and Firefox observations are retained separately: feedback appeared about 0.9/1 ms after the cancellation event was received, while final completion took about 1,027/45 ms. Driver request-to-completion included additional automation overhead. WebKit reached the three-download audit cap before the driver's natural cancellation input arrived; its natural timing is not qualified, and those failed runs remain failures. Preparing focus earlier did not establish a valid natural measurement. The deterministic callback hold resolves coverage of cancellation semantics, not that timing gap.

The controlled run's 50 ms heartbeat observed maximum event-loop gaps of 99/125/186 ms across Chromium/Firefox/WebKit. These are single-run diagnostic gaps, not physical input latency, p95 or an agreed reference-device threshold. The controlled cancellation timings include injected delay and are not used as performance claims.

Retention and applicability

Host: macOS 26.7, Node 24.16.0. All three final journeys passed without retries, skips or timeout changes. The required runner now discovers 186 tests in ten files, including one dense-export case per engine. Discovery is not a full suite run. Future publication budgeting must include this case; no hosted qualification has been initiated.

Session 20233d90-7a79-4423-a564-f75af5b08662 retains:

  • files/dense-export-qualified-local and dense-export-summary.json: source- linked final receipts, allocation/encoding observations and the nine actual downloaded PNGs across the three engines.
  • files/dense-export-initial-chromium: original unacknowledged-wait observation.
  • files/dense-export-final: natural Chromium/Firefox results and the failed WebKit safety-limit run, not an all-engine passing campaign.
  • files/dense-export-webkit-cancel-fix and dense-export-webkit-encoding-cancel: unsuccessful natural-input harness attempts, retained with audits, screenshots and traces.
  • files/dense-export-controlled-webkit: the first successful controlled pending-encoding case, superseded by the final source-linked campaign.

The early failed runs' preliminary downloads were not retained; their audit records and traces identify the bounded requests. The final harness now retains unverified received files on failure before cleanup, and the final passing campaign retains the required first/last/player files. Synthetic fixtures and historical outcomes are unchanged. No cache/profile cleanup was performed.

This closes the bounded F05 page-allocation, selected-file validity, controlled-cancellation, retry and source-preservation gap. It does not establish full-batch completion, native heap/GPU peaks, natural WebKit cancellation latency, cold/offline loading, representative hardware or physical printing. Only export-dialog cancellation feedback changed in application code; unrelated render/gesture campaigns were not invalidated or repeated.

Accounting: zero hosted triggers/reservations against zero approved runner minutes. No push, PR mutation, deployment, workflow run, dependency installation, paid-provider call, agent delegation, native setting change or physical print. Local compute/disk were not metered; account-wide usage remains unknown.

Portable local evidence archive

September 30, 2026. The owner approved preparing a portable evidence archive without uploads or new qualification campaigns. The snapshot is source 98cdada6dc7cd152071cb49b6300e89b1474e6b7; this later handoff documentation does not retroactively change the archive's source or historical results.

Session 20233d90-7a79-4423-a564-f75af5b08662 retains these files under files/portable-evidence-2026-09-30/:

File Purpose
dungeon-mapper-evidence-98cdada-2026-09-30.tar.gz Portable archive, 346,352,032 bytes (330.3 MiB)
dungeon-mapper-evidence-98cdada-2026-09-30.tar.gz.sha256 Whole-archive SHA-256
packaging-result.json Selection counts, archive identity and absent historical references
roundtrip-result.json Actual extraction/integrity results and original-file preservation check
archive-readme.txt, verify.mjs Transfer guidance and dependency-free Node verifier, also inside the archive
package.mjs, check-roundtrip.mjs Task-local assembly and roundtrip scripts, not application tooling

Archive SHA-256: 230fad77514811b361e72b28c00d27e7971faf8b125893ef012b8e2dcdc12203.

The archive contains 2,170 manifest-covered payload files plus the manifest and its checksum. These include 1,644 unchanged original evidence files, 62 raw browser campaign reports, the tracked source/fixtures/licenses/review sheets, and binary-capable commit patches from the included upgrade predecessor 72f32e1 to 98cdada. Existing baseline/candidate/rejected distributions and identities remain distinct. Failed CI, rendering experiments and natural WebKit cancellation attempts remain failures or incomplete observations.

The package's selection.json lists included roots and exclusions. Thirty exact historical artifact-root references in this handoff are not present in this session, including earlier Library, preservation and renderer campaigns. They were not searched for in other sessions, recreated or represented as packaged. This is a portable selection of available evidence, not a complete archive of all project history. Raw reports preserve original paths and source metadata. Six standalone review pages also have derived portable copies; nine absolute file links were rewritten with originals and replacement paths recorded separately.

Persistent browser profiles, dependencies/symlinks, duplicate HTML report viewers, native validator executables, conversation attachments and redundant iterations are excluded. Source licenses and asset notices are preserved. Synthetic private project/session files, traces, local paths and source/author metadata remain in the selected evidence. The archive is not sanitized for public distribution and is not a community map-sharing export.

Fresh extraction into an empty task-owned directory verified all payload hashes. Modified, missing and extra payloads, symlinks, hardlinks and modified manifests were rejected. All 1,644 selected originals still matched after packaging; the twelve dense-export file digests and both eighteen-file baseline/token-candidate manifests matched their retained historical identities. Only temporary assembly/extraction copies were removed after these checks. Original evidence and the final archive remain unchanged.

For an authorized transfer, retain the tarball and its external digest, check the received tarball's SHA-256, extract into a fresh directory, then run node verify.mjs from its root. No dependency installation is needed. Checksums establish byte integrity, not a signature or independent provenance. Record the destination, access policy, retention policy and verification of the received copy before closing durable retention. Current session staging on the same filesystem does not satisfy that requirement.

The owner suggested GitHub as a possible destination and a future optional community-map publishing integration. A repository Release attachment is a candidate home for this evidence archive, not an approved upload. User-owned editable map repositories with previews, explicit licensing and publication review are separate product scope. A GitHub login alone does not authorize application writes; any integrated flow needs explicit access authorization. No repository, Release, authentication service, map-publishing feature, storage policy or public license was created or selected here.

Accounting: zero hosted triggers/reservations against zero approved runner minutes. No push, PR mutation, workflow run, deployment, upload, paid-provider call, agent, application rebuild/test campaign or native-setting change. Archive integrity checks do not replace application or hosted qualification. Local disk/compute were not metered; account-wide usage remains unknown.

October 1 PR #185 check-fix investigation

The accumulated branch was published once at c87cc4b1ef9f58ec1c46042d1ff4ea267e5fe47f. Read-only inspection of run 36913680885 found it completed with failure, not pending. This updates the preceding local-only publication status, but does not turn historical or local evidence into a hosted qualification receipt.

The supplied September 15 Firefox failure is the application focus-clipping defect documented above. Its correction aa40a14 and strengthened regression 0d0e0bb are included in the published head. Its Browser qualification failure was a correct downstream result, not a second application defect. Neither historical failure warrants rerunning the old revision.

The October 1 run failed earlier: Build and test passed lint/build, then reported two failed tests and 1,144 passed. Browser journeys never started. The aggregate correctly rejected RESULT: skipped. Both new failures were introduced by this PR's intended provenance addition in 8eaafa0, not infrastructure or pre-existing main behavior:

  • The full-catalog sample equality fixture omitted the new inherited license record. The local correction explicitly pins its complete title, author, license, URL, notice, version and empty asset-credit list while retaining whole-project equality. It does not discard or ignore provenance.
  • The print companion's launch-sample fingerprint still identified the source immediately before that provenance addition. Inspection confirmed the source diff only added the intended metadata. The local correction records sha256:88dd81c171c83ad7ab5cf374a30940b4d50e8c0621a3dc0fe9c87b3393eb9190. The exact hash assertion and all artwork/legacy-token hashes are unchanged.

Both failures reproduced locally before correction. After correction, all 55 tests in the Folio-token, print-companion, utility/hook provenance, launch-sample and premade-map selections passed on Node 24.16.0. Zero-warning lint, the production build and git diff --check passed. The build retains its existing large-chunk advisory; npm also reports the existing unknown global-config warning. No dependencies, renderer geometry, CI configuration or qualification assertions were weakened. This selection is not a new full-suite or hosted/browser qualification receipt. The previously retained current-head focus tests and Firefox 200% production case were not repeated for these metadata/fixture changes.

Hosted accounting and publication hold: The owner approved 200 aggregate runner minutes for one PR-update trigger and one eventual post-merge CI/Pages trigger, with merge requiring separate authorization. The PR-update trigger has been consumed. The investigation itself initiated no hosted trigger. The job timestamps for the completed run are:

Job UTC start to finish on October 1 Duration Conservative Linux minutes
Build and test, 110542298313 19:21:29 to 19:23:34 125 seconds 3
Browser qualification, 110543160601 19:23:37 to 19:23:40 3 seconds 1
Browser journeys matrix Skipped before runner execution None observed 0
Package qualified Pages Skipped before runner execution None observed 0
Deploy qualified Pages Skipped before runner execution None observed 0

Thus the run accounts for 4 conservatively rounded runner minutes, with no running or queued jobs. The coordinator can release its 90-minute active-run reservation, retain 100 minutes for the conditional post-merge run, and record 96 minutes unallocated within the 200-minute ceiling. These figures are timestamp-based accounting, not an invoice or account-balance lookup.

Publishing this local correction requires renewed scope approval for exactly one additional PR-update trigger; no automatic rerun is proposed. The September 15 representative run used 35 rounded minutes across all five jobs. Given the expanded branch and changed artifact pipeline, estimate 35 to 60 minutes with uncertainty, and propose a 90-minute ceiling for the new trigger, including setup and all three browser legs. If approved, 4 consumed plus 90 for that run plus the existing 100-minute post-merge reservation leaves 6 minutes; this proposal is not approval. The workflow's PR job timeouts sum to 77 minutes before extra runner overhead. Build/browser artifacts retain for 14 days, conditional Pages packaging for one day, and npm caches also consume storage. Actual billed usage, available account allowance and storage charges remain independently unverified. Do not push, rerun, merge or bypass required checks until the coordinator obtains the necessary authorization.

October 1 follow-up: browser run 36914937106

The owner subsequently authorized one additional PR-update trigger within the same 200-minute task ceiling. The provenance correction was published as 96757fedaeb0afcf491216019e5d1ca61eba1a99, consuming that authorization. Run 36914937106 is a new run, not a rerun of either historical failure. Local HEAD, the remote branch and PR #185 all matched that revision, with a clean working tree, at the start of this investigation.

Reported check Classification Evidence and disposition
Browser journeys (chromium), 110547346898 PR-caused test integration omission 61 cases passed; publication/player preview and session/player display failed during fixture import because the generic file-input locator matched four controls. Fix the locator in this PR.
Browser journeys (firefox), 110547346677 PR-caused test integration omission The same two cases failed at the same fixture selectors; 61 cases passed. The cause is shared with Chromium, not the historical focus-clipping defect or transient infrastructure.

The base revision 72f32e16 has one Library file input. This PR added creator package/folder inputs in 4525a58 and a package-comparison input in 5fbd710, but left the two older journey fixtures selecting every input[type="file"]. Playwright correctly rejected the ambiguity before either journey could exercise its behavior. The tool itself is not broken.

The correction changes only the import locators in src/test/ux05Audience.browser.mjs and src/test/ux06Session.browser.mjs to getByLabel('Import project', { exact: true }). Strict single-target selection, native file-change handling, import review, save/reload, player privacy, downloads, session recovery and all downstream assertions remain intact. There is no .first() fallback, skipped case, weakened assertion or change to application code, dependencies, build configuration or CI infrastructure.

Local validation: The unchanged production source built successfully on Node 24.16.0. Before correction, all four selected Chromium/Firefox executions reproduced the exact CI ambiguity. With the two-line correction, both complete journeys passed on Chromium 151.0.7922.34, Firefox 153.0 and WebKit 26.5: 6 passed, zero skipped, zero flaky, in 73.2 seconds. Zero-warning lint and git diff --check also passed. The existing build chunk-size advisory and npm global-config warning remain. This is focused macOS evidence, not a full suite or a replacement for exact-head Linux hosted qualification.

The retained evidence in session 20233d90-7a79-4423-a564-f75af5b08662/files/ uses the prefix pr185-browser-36914937106: sanitized Chromium/Firefox job logs, jobs.json, build.log, baseline.log and baseline/, candidate.log and candidate/, and lint.log. The candidate report identifies the tested source as 96757fedaeb0afcf491216019e5d1ca61eba1a99-with-import-locator-fix. Reports and generated artifacts are not committed.

Bounded hosted snapshot: The saved all-jobs response was observed at 19:49:00 UTC on October 1. The run was still in progress. No waiting or polling loop was used, so the following is not a claim of its final state:

Job UTC start to finish Observed completed duration Rounded Linux minutes
Build and test, 110546471561 19:31:41 to 19:33:46 125 seconds 3
Browser journeys (chromium), 110547346898 19:33:50 to 19:45:04 674 seconds 12
Browser journeys (firefox), 110547346677 19:33:50 to 19:45:49 719 seconds 12
Browser journeys (webkit), 110547346896 Started 19:33:50, still running Not final Reserved, not counted as completed
Browser qualification Not yet present in the job response; dependent work remains Not observed Reserved
Package/deploy Pages Not present; PR events do not satisfy their main-only conditions No execution observed No PR execution expected

Both completed browser jobs passed export/offline/update journeys, unchanged distribution verification and evidence upload. WebKit's floor preflight had passed; its browser suite was still running, with export, verification and upload steps pending. No terminal WebKit or aggregate result is inferred.

Completed jobs in this run consumed 1,518 observed seconds, or 27 minutes when each job is rounded up separately. Adding the previous run's 4 minutes gives 31 known rounded minutes. Of the active run's original 90-minute reservation, 27 is now consumed and 63 remains reserved until completion. Keep the separate 100-minute eventual main/Pages reservation, leaving 6 unallocated: 31 consumed + 63 active-run reservation + 100 post-merge reservation + 6 unallocated = 200. Do not release the unfinished run's remaining reservation based on these partial results. These are timestamp estimates, not an invoice; account balance, billing and artifact/cache storage charges remain unverified. The separate unused fixture-only 24-minute ceiling is not available to this task.

Publication hold: The correction is for a local commit only. This investigation initiated no push, rerun, dispatch, hosted review or merge. There is no authorized third trigger. After refreshing completed-run accounting, the coordinator must obtain explicit approval for one push to evillollive-editing-workflow-qualification, expected to produce one new PR CI run. Estimate 35 to 60 aggregate runner minutes with uncertainty; retain the 90-minute conservative ceiling for all jobs/setup, without assuming a rerun. That ceiling cannot fit within the existing 200-minute total while preserving the 100-minute post-merge reservation and already observed usage, so the total budget or reservations also require an explicit owner decision. The workflow's 77-minute summed PR timeouts, 14-day build/browser artifacts and npm caches are unchanged. No publication, merge or CI bypass is authorized by local success.

Remaining release gates

Subsequent housekeeping removes 68 of the original 73 hook warnings and 19 dependency-rule suppressions from map-state, level-management and history hooks. The September 12 cleanup resolves the other five warnings in App.tsx and MapCanvas.tsx, making the current lint allowance zero. Existing scoped rule suppressions outside that cleanup remain; this is not a claim that every suppression has been retired. History helpers have stable identities; editing callbacks declare current dependencies while preserving per-level undo and stale-project callback rejection. The counts above describe the earlier milestone evidence, not the current warning baseline.

UX-09 remains in progress. These automated journeys cover parts of roadmap section 8, not the entire nine-journey/seven-fixture acceptance matrix. UX-07's initial-release digital art is implemented and owner-approved, with decisions recorded in the art closeout ledger; ART-05/09 are deferred expansions, not launch dependencies. UX-00 participant research remains deferred with zero participants under UX-09C.

The bounded A-DATA preservation and A-EDIT editing journeys are covered above. UX-09A still requires exact-head CI, final-candidate applicability for the retained real-version upgrade/rollback evidence, a measured desktop performance/support scope and exact-revision release handling. Public workflow guidance is locally updated and digital-art owner decisions are complete. Physical printing also blocks UX-09A: digital DPI/layout evidence cannot replace A4/Letter paper measurement, assembled overlap and monochrome readability acceptance. See the release gate ledger and print procedure.

Human screen-reader review (VoiceOver/Safari and NVDA/Firefox), WCAG 2.2 AA evaluation beyond scoped critical-action checks, broader native zoom and physical touch/pen/keyboard behavior, and remaining device-specific performance qualification belong to UX-09B. They are deferred from early access, not passed. Known inaccessible critical actions still block release. Desktop performance scope must be resolved in A; local F05 diagnostics do not establish reference acceptance, and the existing full-release budgets remain unchanged. No latency threshold or usability success rate is claimed from test-run duration. Existing React act and jsdom navigation warnings are not silently presented as a warning-free unit-test baseline.

Rollout and rollback

There is no data schema, serialized project, artwork or service-worker policy change. Keep both required checks strict and up to date on the default branch. Older PR branches must update to include this workflow before merging; a missing required browser check should block, not be bypassed.

Revert a problematic product change through a reviewed PR while retaining the qualification jobs. A full rollback to a pre-UX-09 workflow would stop emitting Browser qualification; restoring that workflow and changing the server-side required-check policy must therefore be a deliberate coordinated operation, not an emergency silent weakening. Preserve project backups and follow UX-08's explicit saved-workspace update process for any deployed rollback.

The A-RELEASE pipeline is now implemented locally, not remotely qualified. The remote baseline still has the independent Pages workflow until budgeted publication and merge. Require successful qualification of the published revision, resolve final-version rollback/recovery applicability, and preserve a versioned release evidence/limitations summary beyond the current fourteen-day CI retention.

September 28 design-only follow-up: The exact-revision Pages design records the proposed same-workflow dependency graph, canonical artifact reuse, explicit integrity checks, main-only deployment, superseded-job versus active-deployment handling, rollout/rollback and retention decisions. Read-only inspection confirmed both strict required checks and the main-only Pages environment; neither changed. The design's acceptance table is planned work, not simulated or hosted passes. The subsequent core implementation was explicitly approved and is recorded above. Publication and hosted verification still need a separate numeric budget. The original design task initiated zero hosted triggers and made no workflow/environment or remote changes.