Bind phase trustees and verify native ACL fixture cleanup

[CompleteDotTech]

The phase-trustee tests and installed checks need the same source and fixture closure in source archives, sdists and wheels. Windows qualification also needs a genuine restricted-primary positive, and diagnostics must report failed restoration for every attempted fixture.

This change packages that closure, compares numeric trustees using native SID identities, retains the production strict guard, and adds the native qualification gates. Diagnostic cleanup now fails when any captured fixture cannot verify restoration, while keeping observationally inexact unprotected installation separate from failed cleanup. Failed Job assignment directly terminates the unassigned owned child; null or incomplete child reports cannot become successful evidence.

Refs #110. Refs #94 for the six-case native converter-parent gate; genuine private Docling acceptance remains incomplete.

Validation of signed candidate a23101e56c25289c7c64858b63ca68efdfaa7b89:

• Independent actual Astra source and complete-package review, followed by evidence-caller rereview: READY_FOR_SIGNED_NATIVE_CI; selective restoration and resource-caller findings are resolved. The signed source blobs exactly match the reviewed candidate. Final signed-head approval remains required.
• Fresh Linux Python 3.12.14 full validation: 1,792 tests, 1,669 passed, 123 skipped, zero failures or errors; producer exit 0. Generated evidence was preserved and all 720 tracked files restored.
• Fresh hosted checks on this signed head completed with six Linux jobs passing and four Windows jobs failing across the normal push and pull-request matrices. Windows 3.13 reports completed: false for a restricted-primary helper child with exit 0 and verified cleanup; Windows 3.12/3.13 also expose a v1 refusal probe whose DACL restoration differs from its pre-state. Source and extracted-package failures are preserved and repairs are in progress. This head does not satisfy native acceptance.
• Focused Python 3.11/3.12/3.13 checks: 39 zero-exit commands. The helper has 155 tests, with 16 Windows-only skips on Linux. Native qualification requires bare OK without skips.
• Targeted modeled mutations: 19 detections with one unmatched pattern preserved; its corrected replacement was detected separately. These checks do not establish comprehensive or native coverage.
• Current source ZIP: 719 members including a 718-entry manifest. Supported setuptools 68 and latest sdists have 223 files; both wheels have 105 members, including 99 package files. Source, sdist, wheel bytes and wheel RECORD entries were independently audited. The existing packager exclusion of docs/workflow/render.mjs remains explicit.
• All 28 extracted-package commands passed with expected counts. Three actual isolated wheel installs passed portable checks outside the checkout; wrong-digest and required-Windows negatives returned 1 as intended. Third-party dependencies were copied offline from matched prior environments; the current wheels were installed with pip.
• Owned validation wrappers now guard explicit workload and utility launches with the same argv, working directory and redirections that they execute. Prospective evidence includes 20 resource-witness tests, 19 launcher tests, 160 matched launches across six wrapper paths and 29 defer/lock-loss cases. The path proof uses isolated stubs and does not establish package or native acceptance. Historical resource gaps and failed attempts remain preserved; descendants of a guarded command inherit its lock rather than receiving separate witnesses.

The public versioned method-map dossier contains the complete 87-entry baseline comparison and one changed phase_authority.py entry. It does not rebind private pilot receipts or establish empirical qualification.

Before merge: repair the native failures, refresh affected validation and packaging, verify the resulting signed head, obtain genuine Windows 3.12/3.13 source/extracted/installed positive and negative evidence and all exact-head CI, obtain actual Astra approval of that exact final signed head, then use normal merge gates and verify post-merge CI/readback. Linux passes and the bundled historical Windows report do not satisfy those gates.

Unpublished repair status (2026-10-09): actual Sonnet 5.5 completed the repair handoff after capacity was restored on the same account and model. A separate actual Astra review completed full reading of the 19-file source/manifest candidate and all 55 recursive evidence-code files, independently reconciled the retained results, and returned CHANGES_REQUIRED with six remaining gaps: complete dependency/mutation inventories, complete traceback-frame/message validation, ACE-derived component facts, authoritative method-report identity, all-five thread observations for each child, and actual-main coverage of the nested retry failure. No source/caller readiness or approval has been issued. The exact repair tasks are prepared; the same-account Sonnet session limit currently reports a reset at 03:10 UTC on October 9.

The current focused phase completed 39 module/interpreter runs across Python 3.11/3.12/3.13 and 39 targeted modeled mutation checks, including restoration and the final baseline. Helper runs report 216 tests with 17 Windows-only skips on Linux. The final controlled caller proof completed with producer exit 0, 217 scenarios and no reported scenario or global failures; retained caller-code inventory and bytes match the current candidate. An earlier 217-scenario proof failed four cases and remains preserved (213 passes). These are portable and controlled-proof results, not native Windows or complete package acceptance.

The repairs address nested privilege-retry token-close reporting and retained Job-handle ownership after failed setup cleanup, together with required-child inventories, summary grammar, terminal assertion classification, installed-report schemas, method-map identity and caller-proof bindings. Independent review must confirm these repairs and their evidence. Exact DACL restoration remains mandatory, observational control-flag drift remains diagnostic, and the production strict guard is unchanged.

The earlier 1,808-test full validation and package evidence predate these repairs and remain historical. The current manifest is deliberately stale and must be regenerated and audited after source/caller readiness. Failed, deferred and interrupted attempts remain recorded. No current native Windows repair qualification exists.

This PR remains draft at signed head a23101e56c25289c7c64858b63ca68efdfaa7b89, with six Linux successes and four Windows failures on its normal hosted matrices. Before merge: complete independent source/caller review, run fresh complete validation and genuine current package/installed checks, publish signed changes, obtain genuine Windows 3.12/3.13 source/extracted/installed qualification and all exact-head CI, obtain actual Astra approval of the exact final signed head, then use normal merge gates and verify post-merge CI/readback. These repair results do not establish private empirical acceptance.

[CompleteDotTech]

Draft delivery checkpoint for published head b5696985b49d7c25c7540d435b0b8056cfe3a5de.

Independent actual Astra review of the current unpublished repair returned CHANGES_REQUIRED. Two source fixes remain before successor publication: terminating the owned suspended child when Job assignment fails, and rejecting missing or non-object child reports as incomplete evidence. No final-head approval or successor publication is claimed.

Portable focused checks and causal mutation evidence are retained. Fresh full validation was interrupted without a successful exit receipt, so it is not acceptance. The published head's native Windows checks remain failed. The current unpublished source archive matches its manifest, while fresh distribution and installed-package qualification remain unfinished.

Next: Sonnet repairs, affected validation and refreshed packages; then a signed successor with Windows 3.12/3.13 source, extracted-sdist and installed-wheel positive/negative checks, independent Astra review of the exact final head, normal merge gates and post-merge verification. Existing converter-parent native regressions will also be checked separately from genuine Docling smoke qualification.

Refs #110, #94. Historical pilot authority and empirical qualification remain separate.