# P5-15-04 — Targeted acceptance repairs

This successor retains the [02 failed observations](../p5-15-02/DEFECTS.md). The owner subsequently authorised continuous, bounded repairs before completing document acceptance. Each demonstrated failure was reproduced and corrected independently; production `main` remains unchanged and NO-GO.

## Completed repairs

| Finding | Demonstrated cause and smallest correction | Native/regression proof | Commit / PRES |
|---|---|---|---|
| A11Y-01, P1 | Nonnative file-input labels did not activate the picker on Enter/Space. Existing file controls now use native buttons and the same input click/validation path. | Eight added frontend cases; six native keyboard and three pointer chooser outcomes; no accepted-file/import/ownership change. [Record](file-picker/README.md) | `57054349ef7816b84a31c29a8c847a734ab22479`; W08/J10 |
| A11Y-11, P1 | Initially hidden compact-mode preview pages published ready after measuring zero width. Pagination now awaits actual usable layout through one ResizeObserver before the unchanged physical calculation. | Two new hidden-initial fixtures, one/four pages; reveal/resize/scale lifecycle, existing mobile native PDF and both unchanged fidelity corpora. [Record](mobile-preview/README.md) | `c22f3fc753f48d17c09e1cccd0b694957905c77c`; C06/C09/C12 |
| A11Y-02, P2 | MUI attempted autofocus while the transitioning drawer was visibility-hidden. Settled entry now focuses the existing first item when focus is on the dialog or outside it. | Nonmutating focus trace; reduced-motion Enter/Escape/trigger return and emulated rotation. Modal protection remains enabled. [Record](mobile-drawer/README.md) | `b377f0f62f661bb7a63f1a2b784f22e8b9b52de4`; E07 |
| A11Y-03, P2 | The CV source link used the API path `/career/profile`, which is not a frontend route. It now uses existing `/career?section=profile`. | Native exact destination, canonical form and Back-to-CV assertion; no second source editor or mutation. [Record](source-navigation/README.md) | `b5bda1caa58749c0171a83101fb4e3bf337c23e7`; C07 |
| A11Y-04, P2 | Standalone report links had an 18px hit height. The existing shared link treatment now has 24px minimum width/height. | Native EN/light phone and NB/dark tablet bounds, keyboard activation and exact custom-status filter/job; report values unchanged. [Record](report-targets/README.md) | `22455bba0a016f0cefabde46c539df588f789667`; S01/J05 |

No API/schema migration, destructive operation, renderer geometry/font/print-rule change, provider activation or production deployment was required. The last application change is `22455bba0a016f0cefabde46c539df588f789667`. All code repairs and their focused evidence were committed/pushed immediately. The [complete post-change regression](../p5-15-06/README.md) passes fresh:1,033backend/498frontend/148browser/3diagnostics and unchanged revision/fidelity/security gates. Scoped repairs are GO; complete affected human/device/deployed acceptance remains conditional.

## Reliability and capacity investigation

The diagnostic qualification extends the existing native preference scenario to twelve alternating English/Bokmål language and notification writes while four real report/notification reads overlap. All twelve rounds returned language204/notification200, all four reads200 and correct owned readback, with EN/enabled-reminder persistence after reload. This directly exercises the recorded overlap without mocking persistence or inventing a common cause for two historical failures. Historical500 causes are **not reproduced in this finite run**, not repaired or closed. [Actual evidence](evidence/reliability-capacity.json).

The new Board diagnostic created synthetic owned populations of 200 and 1,000 jobs and passed exact real API/DOM ID membership at both sizes. Observations:200jobs/API33ms/render1,040ms/236,253bytes;1,000jobs/API19ms/render1,612ms/1,182,578bytes. These are finite local measurements, with no approved production performance threshold to declare satisfied. Neither a query cap nor speculative pagination/virtualisation was introduced. The existing full-array O(N) architecture and physical-touch acceptance remain open. All three separate diagnostics pass, versus the inherited two; extending preference rounds does not add twelve test cases. The diagnostic/evidence unit is committed/pushed independently.

The [CV save-state investigation](../p5-15-01/INTERMITTENT_FAILURES.md) qualifies the editor's existing 800ms autosave boundary with a controlled test clock: pending draft/focus/no write before it, exact versioned save afterward. No saving implementation changes. The untraced historical incident remains OPEN; passing reruns alone do not establish its original cause.

## Acceptance disposition

[Remaining findings](DEBT_DISPOSITION.md) distinguish reproducible functional repairs, contextual rule warnings, unavailable human/device checks and deployment dependencies. There is no newly demonstrated security/data-integrity failure in these repairs. The whole-group reconciliation is **0 PASS / 56 PARTIAL / 0 FAIL / 1 DEFERRED =57**; five scoped fixes do not confer complete accessibility or whole-feature acceptance.

P5-13 remains deferred/unqualified/Production-withheld, with 0/366 configured calls, zero human semantic reviews and no review UI. Quick preview and `followMode` compose UI remain approved separate work. Mandatory human screen-reader acceptance is postponed at the owner's request and NOT COMPLETE, not waived. Production ingress/recovery/configuration/schema/rollback and Linux rendering qualifications remain independently open.