cesnimdaP9-05 — Documents, edge cases and finite performance
The final separate Phase9 diagnostic lane completed 18/18, zero failures/skips/flaky/retries, on the frozen825-input manifest (phase9-frozen, exit0 at20:25:16UTC). It includes12 route/theme/locale/width cases, four settled accessibility cases, one RD21 characterization and one bounded workload. A passing characterization does not qualify the unsafe behavior it records. The earlier18-case run and the later16pass/2fail transition observation remain superseded raw evidence.
That diagnostic snapshot is retained as source-manifest-before-admin-fixture.json. The final ordinary-run snapshot changes only the locale-sensitive Admin test locator; exact delta proves product/build/diagnostic/core inputs are unchanged. No diagnostic result is claimed to execute an ordinary fixture it does not load.
Native documents
Fresh real preview and native exported PDF bytes from the retained corpus passed 16/16 fixtures /36 pages under the unchanged oracle: exact page count/membership, physical dimensions ±0.5mm, marker positions ±2mm, font sizes ±0.5pt and original RGB tolerance. Nine defaults, long EN/NB, 8mm Letter, 28mm roomy, grid/bubbles, long sidebar and unbreakable boundary fixtures are included. No renderer geometry or tolerance changed.
Portable oracle results retain all measurements. The final frozen ordinary run's native corpus producer passed; the unchanged oracle again passed16/16 over its fresh36 pages. All36 regenerated page thumbnails were actually inspected in sheet1, sheet2 and sheet3, plus the full-size preview/PDF unbreakable-and-boundary/comparison-4.png. Inspection found no visible clipping/overlap/blank output in those views. Several last pages contain little content; that is recorded layout behavior, not missing text according to the marker oracle. Thumbnail/spot inspection is not print/readability, photo/crop, every glyph or intended Linux-font acceptance. Whole-suite completion remains a separate gate.
Existing browser fixtures also cover hidden preview mounting, image-layout readiness, dirty edits, stale CV revisions, preview transport recovery, canonical versus variant overrides, frozen submitted packages/archived binaries and missing optional sections. Their full-suite outcomes belong to final regression. Native PDF admission uses inert child processes and is not added to fidelity counts.
The separate Settings corpus passed the same unchanged oracle over 8 fixtures /24 pages: EN/NB × light/dark, before and after actual document-control changes. All24 page thumbnails were inspected in Settings sheet1 and sheet2; a full-size light-EN before preview/PDF comparison was also inspected. No visible clipping/overlap/blank output was found in those views. Measured results retain the before/after page dimensions, marker/font/colour agreement. These are separate24 pages, not extra fixtures in the16/36 baseline corpus, and do not close human/Linux/print requirements.
Sustained local workload
Raw finite measurements: 65.303s, two paced readers, 240 successful reads, 250 owned applications, eight Career roles with48 distinct long bullets, and20 native navigations across Dashboard/Table/Board/long CV. The same saved canonical profile and250-job membership remained intact afterward. CV readiness waits for settled pagination and verifies the final bullet's page membership and multiple rendered pages before recording timing.
API latency across these reads: p50 23ms /p95 76ms /max115ms. Native ready times in this run: Dashboard1377–1728ms; Table995–1249ms; Board1530–2238ms; long CV2038–2310ms. These include browser boot/auth/render work and automation overhead. The host reports Windows,16 logical CPUs,34,145,554,432 bytes RAM and Chromium151.0.7922.34; API uses isolated SQLite and frontend uses the rebuilt static export. This is a shared development host; final Jest overlapped most diagnostic collection and completed during the sustained case. Timing is not an isolated machine comparison.
CDP CPU/task/heap/DOM metrics are retained for every navigation. They are observations, not leak diagnosis or attributed backend resource profiling. Two paced readers for a minute do not establish capacity/SLA, sustained drag frames, physical mobile performance, long-term memory stability, MariaDB/Linux behavior or real model inference throughput. No production load was generated.
Retained ordinary fixtures extend edge-case coverage to no-data,1,000-job Board membership/filtering, long timelines/profiles, revision conflicts, stale sessions, HTTP/network failures, missing/foreign references and recoverable drafts. Remaining worker durability, creation idempotency and high RD21 stale writes are explicit risks. No synthetic data was loaded into production.
A separately isolated volume supplement adds3 passing native API cases:100 CV revisions plus additive restore101,100 long manual timeline messages with exact membership and foreign-owner denial, and100 cancellation notices with bounded listing/read/dismissal counts. Its source hashes and results are separate from the825-input ordinary-run snapshot. This establishes finite persisted membership/count/revision behavior, not large-history browser capacity.