- Decision
- ADR-091
- Date
- 2026-09-18
- Status
- InIn progress, not closed. Tasks 1-6 complete and independently verified; Tasks 7 (hardware build), 8 (hardware verification), 9 (real closing ADR) remain open.
- Mentioned by
- ADR-092
1Decision
Interim record for docs/superpowers/plans/2026-09-17-duress-unlock-final-closure.md's Tracks 1-2 (Tasks 1-6 of 9): audit and squash three confirmed dead-end patch sequences, and make two disclosed completeness gaps (RealUserQuiescer's dead failure path, UserRepositoryImpl's stale-cache bug) genuinely real. This is not the plan's own closing ADR — Task 9 owns that, once Tracks 3's build and hardware sweep (Tasks 7-8) actually run.
2Why and evidence
Task 1 audited all 104 patches across four repositories against the real branch tip (not from memory), confirming three dead-end sequences and correcting the plan's own stale framing on two points: frameworks-base/0024 bundles four distinct changes across two files, not the plan's stated two in one (a USER_SETUP_COMPLETE/DecoyProfileState-ordering fix the plan never mentioned had to be identified and preserved separately), and android.multiuser.credential_capture's own upstream default already resolves to state: DISABLED/permission: READ_ONLY on this product's cp2a release config — traced through the full six-hop release-config inheritance chain, not assumed.
Tasks 2-4 squashed the three candidates after per-case owner sign-off (build-release 0001-0004 → one commit 0ded9b0e3972, kept explicit rather than deleted for defense-in-depth, matching this project's own established pattern of never relying on an implicit upstream default for this exact flag; frameworks-base 0018+0024's dead resource-revert hunk → dropped, 0024's three real fixes preserved unchanged, commit 949164d7ae30; frameworks-base 0015+0016 → one commit babea734ee48, landed by the owner directly after this session's harness twice declined to run the equivalent history-rewrite programmatically, real tip 08ba5b75a24d).
Each squash proved behavioral identity before landing — Task 4's proof was the strongest of the three, an exact git blob-hash match (4aeb1e0b71daad2654dc68629d325a4ffcf6f5a1) on a disposable throwaway branch before the real branch was ever touched.
Tasks 5-6 landed two real fixes: RealUserQuiescer.quiesce(int): boolean now returns false on any single process freeze failure (the owner's exact threshold, not a looser "all failed" substitute) and ActivityManagerService.quiesceUserApps is a genuine pass-through instead of a hardcoded true (commit 16e532d86674); UserRepositoryImpl._userSwitcherSettings now merges KeyguardRepository.isKeyguardShowing into its upstream Flow so a keyguard transition re-evaluates it without requiring a Settings.Global write (commit fee7a84087f9).
Both changes verified correct by direct source read (not only trusted from a report) and independently task-reviewed clean.
3Consequences
frameworks-base's patch count is now 87 (was 86 at plan start: -1 net from Task 3/4's squashes, +3 from Tasks 5-6's real fixes); build-release is 1 (was 4); combined four-repository total is 102 (was 104).
Neither Task 5 nor Task 6's new test has ever been executed via real atest — a confirmed, reproducible build-server capacity limit (soong_build's module-graph analysis independently observed at 94% of this VM's 31Gi total memory before being killed, exit=137) blocked every attempt across both tasks, and a third, independent task (an m services identity-proof build) hit the identical signature the same day.
Both task reviewers judged the gap acceptable to proceed past given the triviality and direct traceability of each actual diff, on the explicit condition that it stays open and gets closed before Task 9's real closure — not silently dropped.
A real near-miss was caught and fixed before commit, not after: regenerating frameworks-base's full patch series resurrected the real build server's raw, unsanitized git author identity (the actual LAN account and hostname, not the safe AegisOS Build <build@aegisos.local> placeholder convention) into patches 0001/0002 specifically, the two oldest commits in the series (predating that convention's introduction), whose previously-committed local export had already been sanitized by an earlier session.
Caught by this task's own routine pre-commit grep, confirmed via git show HEAD:<path> that the actually-committed history was never affected, fixed before staging.
4Review when
Before Task 9 writes the plan's real closing ADR: (1) obtain real atest evidence for RealUserQuiescerTest and UserRepositoryImplTest — Task 7's consolidated build will force a real compile of both but that is not equivalent to a real assertion pass; (2) resolve or at least account for the build-server memory-capacity finding before Task 7's own (larger) build attempts the same class of Soong analysis, since three independent tasks hit the identical exit=137 signature the same day; (3) any future full regeneration of frameworks-base's patch series must explicitly re-check 0001/0002's author line — a blanket fresh export will resurrect the same real-identity leak this entry describes catching, every time, until those two commits' own git history is separately addressed.
Read from docs/governance/decision-log.md, entry ADR-091.