docs: publish Phase 9 requalification decisions and evidence

2026-10-24 18:32:24 +02:00
parent 7092697cb3
commit 3727a723e8
22 changed files with 53642 additions and 42 deletions
@@ -1,7 +1,8 @@
 # Retained attempts and qualification-harness repairs

-Application source and rebuilt artifacts remain at the security-pin checkpoint `780b9bc`.
-The subsequent repairs below affect qualification tooling only. Historical acceptance reports
+Application source initially used security-pin checkpoint `780b9bc`; the subsequently reproduced
+document-index provider defect is repaired in `2b26953` with its own fresh full regression.
+The fixture repairs below affect qualification tooling only. Historical acceptance reports
 are unchanged. Successor counts exclude earlier attempts; no failed startup or short run becomes
 a passing database, workload or PDF qualification.

@@ -13,6 +14,7 @@ a passing database, workload or PDF qualification.
 | Schema ownership before diagnostic write | Independent review identified successful CREATE followed by a failing log write could leave cleanup ownership false | Record successful CREATE ownership before writing its log. Exact old-checkpoint source replay fails the injected-error regression; corrected source passes both that case and failed-CREATE/no-DROP authority. This is offline harness proof, not an executed data-loss event. |
 | Initial sustained workload setup |250 synthetic applications were created; stored canonical Career JSON omits null object properties while API DTOs include them. Exact cross-format comparison failed before preflight/timer; zero timed reads/navigations; schema cleanup succeeded | The fixture compares canonical JSON using the existing `StructuredCvProfileJson.SerializePersisted` null-property omission. Full API profile assertions and exact SQL before/after equality stay unchanged; SQL/API revision equality is added. CV override null-as-inherit behavior is untouched. Failed evidence and predecessor manifests remain under `attempts/performance-canonical-wire-failure/`. No15-minute result is inferred from it. |
 | Linux runner error paths, before service execution | Review found post-browser SQL failure could retain evidence exit0, and failed log capture could suppress container removal | Browser and overall exits are separate; any post-check/cleanup error fails the runner; removal is independent of diagnostic capture; unconfirmed API removal prevents schema DROP. Two offline injected-error tests pass. No earlier Linux runtime PASS is claimed. |
+| MariaDB workload preflight after serializer correction | `/api/documents` returned500: illegal mix of collations in UNION; zero timed requests/navigations; schema cleanup succeeded | [Actual provider repair](DOCUMENTS_PROVIDER_REPAIR.md), three real RED/GREEN cases, then a new full35-case MariaDB/backend/frontend/browser checkpoint. Entire previous regression and failed preflight retained under `attempts/pre-documents-provider-fix/`; none is added to successor totals. |

 The schema-ownership regression itself was debugged with offline fixture corrections for the
 temporary evidence root and Docker inspect's list envelope. The retained old-source replay then
@@ -23,10 +25,15 @@ Initial and successor input manifests are retained. The first successor changes
 performance web-server config among the1042 frozen runtime inputs, plus its affected wrappers;
 the next changes only the two performance fixture sources. No API, UI source, built export,
 Release assembly, original PDF oracle or historical test assertion changes during these repairs.
-Each affected lane is rerun against a fresh generated schema. Final accepted results and hashes
+Each affected MariaDB lane is rerun against a fresh generated schema; the subsequent application
+provider repair also triggers a complete fresh regression and rebuild. Final accepted results and hashes
 belong to [validation.json](validation.json); this history does not waive a failing final gate.

 Hosted843/844 retain their original missing-Docker failures. Hosted847 and849 retain expected
 first-step isolation failures, with no MariaDB test execution. [Local Desktop socket recovery](LOCAL_DOCKER_RECOVERY.md)
 is a separate workstation event. Neither its successful startup nor a list-only browser discovery
 report establishes application acceptance.
+
+## Final Linux attempts and fixture successor
+
+`attempts/linux-timestamp-baseline/` retains the original1/5 Linux producer result, source/helper manifests and predecessor Windows settings evidence. Four cases differed solely because POST timestamp precision exceeded durable MariaDB datetime(6). At7092697 the fixture reads the complete durable sibling before actions, then retains strict complete-body equality afterward. Windows4/4 and Linux5/5 successors pass; they are never added to whole-suite362. The unchanged Linux output oracle then genuinely fails11/24 fixtures (65 versus60 pages). See LINUX_PDF.md; no tolerance, baseline, API or rendering contract was changed to obtain a pass.