# P10-ENTRY-FINAL acceptance and handoff

**Safely actionable PDF work implemented; evidence package READY. Phase10 development NO-GO, automatic execution false, production NO-GO.** No protected owner decision is supplied by this report.

Branch `test/pre-phase10-ux-acceptance`, starting accepted package26d531cc9079be3be7601f42cb101035de022a4d. Application correction154a9e89f565e786d7df00bf93d873da900a7c7d; diagnostic checkpoints4024a89/38d7133; handoff source guard52fb91e. Final exact published candidate and independently resolved envelope are specified by [ENTRY_PROPOSAL.json](../automation/handoffs/pre-phase10/ENTRY_PROPOSAL.json) and ACCEPTANCE.json. These later source snapshots preserve all tested application/diagnostic bytes. Existing CV-UX4a0af62 ancestry is retained. No main merge, deployment, model/cache operation, separate-controller mutation or Phase10 implementation occurred.

The mission is complete within its evidence-only authority. The remaining PDF metric decision, human checks and protected entry/runtime/scope actions require the real owner/operator; they cannot be completed by assigning a model-written GO.

## Findings and smallest correction

All48 original revised S01 failures have a specific verified explanation, with all72 passing/failing controls listed in [S01_CLASSIFICATION.csv](S01_CLASSIFICATION.csv) and exact native hashes in [S01_CLASSIFICATION.json](S01_CLASSIFICATION.json). The old test compares a DOM Range top with a PDF extractor's descent-derived one-em cell top, rather than like-for-like geometry. Actual native text matrices, an inline DOM baseline sentinel, independent canvas ascent and bounded SFNT tables reproduce the difference. The cell formula residual is below0.01px. **Every failed case is INCORRECT_TEST_ASSUMPTION with native font-descriptor metadata as a secondary cause.** The Windows-only technical/nordic failures account for the platform count difference. No failure was silently waived.

After the correction, unchanged original S01 results remain **Windows8/36PASS,28FAIL; Linux16/36PASS,20FAIL**. Actual baseline deltas are at most **0.227378mm Windows /0.367938mm Linux** against the same2mm bound; actual S01 ink differences are at most2px/2px. This is a supplementary diagnostic, **not an approved replacement gate**. No golden, font, original oracle or tolerance changed. The candidate metric requires an explicit owner reference/criterion disposition. Windows native/preview Chromium151 and Linux native156/preview151 differ in both OS and release; selection cannot be attributed exclusively to OS without a matched-version experiment.

A separate real defect was found: full names could paint accents above the first print fragment even when preview and PDF agreed. The new regression failed24/36 derived CV variants and2/18 letter variants before the fix. Shared revised CV names and letter sender headings now receive2mm top padding, preserving their existing leading. Actual API/native assertions were RED then GREEN; all derived36CV/18letter clearance variants and actual9CV+9letter European-name controls pass per platform. Version1 markup is unchanged; existing stored/historical revisions are not rewritten. Photo-free themes remain intentionally photo-free. [DIAGNOSIS.md](DIAGNOSIS.md) retains the full investigation and distinguishes superseded line-height experiments from the implemented padding correction.

## Fresh validation after the application correction

| Lane | Outcome and scope |
|---|---|
| Backend |1448PASS,0fail/skip |
| Frontend |1031PASS/112suites,0fail/pending |
| Affected Windows browser |51PASS: focused2 plus disjoint49, exact discovery membership;0unexpected/flaky/skips/retries |
| Additional browser lanes |PDF admission1, RD216, real import8(3+4+1), real MariaDB3(1+2), native Linux8PASS; separately scoped, not added to the51 |
| MariaDB provider |90PASS; transaction/concurrency/rollback/recovery checks; exact owned test cleanup |
| Python |83offline+14real DoclingPASS; model loading disabled; no new inference |
| Baseline diagnostic tests |5PASS; reject actual shifts beyond2mm and wrong metric/sign/unit correspondence |
| TypeScript/lint/build |TypeScript, lint, frontend/API build, isolated Linux image, Windows launcher restoration exit0 |
| Original PDFs, each platform |24CV/60pages and4letters/14pagesPASS |
| Revised PDFs, each platform |36CV/122pages content/bounds/name wrapping/inkPASS; full-height sidebars43/43PASS; matching18letters/38pagesPASS |
| Additional actual native controls, each platform |9no-photo European-name CVs/9pages +9European-name letters/9pagesPASS; real API exports, semantic/bounds/page checks |
| Revised original S01 gate |FAIL28Windows/20Linux, retained unchanged |
| Handoff tests |Final combined19 source/parser/schema/authority/tampering tests; bootstrap7 and old18 are overlapping, never additive. See current indexed native report |

Raw command exits, native membership, before/after source hashes, exact image/DLL/runtime and cleanup are in [TEST_EVIDENCE.json](../automation/handoffs/pre-phase10/TEST_EVIDENCE.json). Frozen application/diagnostic bytes are154a9e8; the focused2 executed the exact later-committed dirty38d7133 bytes, proved by indexed fingerprints. Later runs are at154a9e8 or documentation-only52fb91e. Documentation-envelope binding is not a claim that tests executed at a future commit.

The preceding accepted380-case broad ordinary suite,53 specialized checks and independent helpers are **historical5032ccb evidence**, preserved in Git26d531c and the indexed old manifest; they were not rerun as current380. The current task reran all1448backend/1031frontend tests and the affected51 browser cases plus the listed separate lanes. No old run or repeated native measurement is added to new totals. No new whole-repository security scan, model semantic benchmark, hosted CI, production or human acceptance is claimed.

## Native owner review

[Review guide](REVIEW_GUIDE.md) links eight bounded ZIPs and navigation sheets containing **326 byte-preserved PDF archive entries**:200fresh after-fix documents,108actual before-fix documents,18clearly historical version1 reference entries (the same nine-theme historical files included in both platform bundles). Both platforms include original/revised, short/long/dense/multipage, Bokmål/European diacritics, photo/no-photo, all nine themes, sidebars and matching letters. PDF source/archive SHA256 correspondences are retained. Preview/native before/after name crops aid inspection; JPEGs are navigation only, never goldens or human approval.

The normal local development backend still reports caa0c6a2efea5a6e9d8910f55fb00fd982cd48fd; it was not replaced. Review candidate PDFs immediately, or use an isolated build of the published candidate for interactive confirmation. Do not mistake that earlier runtime for the candidate fix.

## Failures, recovery and limitations

- New assertions failed before the correction; the first GREEN attempt also caught an incorrect new photo expectation for themes whose DTO explicitly says none. It was corrected using the existing template contract, with no product/theme behavior weakened.
- The first post-fix import lane ran concurrently with Linux publish, which removed the Windows apphost; it stopped before cases executed. Owned parser resources were removed. The recovered serial lane restores the apphost before a fresh import/MariaDB run. Original logs/cleanup remain indexed; the failed attempt contributes0passing cases.
- An optional direct native legacy-control CLI timed out before any document completed. Exact root process/profile ownership was checked, its process tree terminated, absence confirmed and its exact temp profile removed. This attempt is not a successful product qualification. Fresh original28/platform plus existing36version1 markup/CSS compatibility tests remain the current compatibility evidence; nine extra native v1 references are labelled historical.
- Human visual/device/AT/copy/physical print acceptance remains NOT TESTED. Revised S01 gate remains FAIL pending explicit owner metric disposition. The finite automated clipping fix does not approve general typography or printing.
- Existing three CV-model probes remain rejected historical evidence; no new inference was run. CV adapter remains disabled; no source-grounded AI accuracy PASS. Dependency/advisory history is carried separately with no new risk waiver: Python vendor Torch404 incomplete, full frontend5low/production0, MySqlConnector moderate, hosted/native-image/operator concerns unchanged.

## Independent entry and release decisions

Actual accent clipping: PASS_AUTOMATED_FINITE_CORPUS. Original PDF compatibility: PASS. Deterministic baseline diagnostic: PASS_DIAGNOSTIC_ONLY. Revised unchanged automated fidelity: FAIL. PDF spacing: PARTIAL. Human/printing: NOT TESTED. CV AI: BLOCKED. Package integrity: independently verified separately from entry. Phase10: NO-GO. Automatic execution: false. Production: NO-GO.

Required owner actions are concrete in [OWNER_REVIEW.md](OWNER_REVIEW.md). First inspect/disposition the revised S01 metric and actual rendered samples; then independently approve the exact published baseline and a proven candidate or assessment-only scope, and qualify the separate real runtime/key/sign-in/composite adapter/policy. No deletion candidate has been proved unused and none is approved. Five conditional-development debt requests retain their source-specific authority; none is granted or executable here, and all controller checkpoint deferrableGates remain empty. Production-only backup/filesystem/recovery/TLS/native-image/manual/provider matters remain independent blockers.

The separate committed controller140e182 remains only an external read-only snapshot:69Python(includes48)+6signature and native MXC canaries are its history, not our tests. Real owner key/sign-in/adapter/activation remain unqualified. No queue, signed approval, policy or other worktree was modified. All57preservation criteria and34routes remain56PARTIAL/1DEFERRED. No cleanup, Product Review, Global Jobs, Phase10 or other initiative was started.