ObliviOS Developer portal Project site

Home / Decisions / ADR-092

Close docs/superpowers/plans/2026-09-17-duress-unlock-final-closure.md.

Closed, recorded 20 September 2026.

Decision
ADR-092
Date
2026-09-20
Status
ClosedClosed. All 9 tasks complete with real evidence; atest for RealUserQuiescerTest/UserRepositoryImplTest remains genuinely UNEXECUTED (no adb-reachable device existed on the build server; not silently dropped, see Consequences).
Mentions
ADR-061, ADR-072, ADR-082, ADR-088, ADR-090, ADR-091
Mentioned by
ADR-094, ADR-095, ADR-096, ADR-098

1Decision

Close docs/superpowers/plans/2026-09-17-duress-unlock-final-closure.md: the branch is feature-complete and hardware-verified. Corrects ADR-091's own citation error for Task 3's squash commit hash (recorded as 949164d7ae30, which is actually an unrelated commit — the real Task 3 squash commit is 64187dc103b5, confirmed by git merge-base --is-ancestor and by the resource value reading false in the current tree).

2Why and evidence

Task 7 (build) required two additional real fixes beyond Tasks 1-6, neither anticipated by ADR-091: 5c837bb3c965 migrated ReferenceGlobalRootComponent.java to Kotlin (KSP, the module's annotation processor since an October-2025 upstream commit, cannot process plain .java @Component sources — masked for six weeks by stale incremental Soong cache); e6f6b9363b83 then f3f3c9ebc980 fixed a real Dagger dependency cycle Task 6's own KeyguardRepository injection introduced into UserRepositoryImpl — the first attempt (e6f6b9363b83, wrapping in dagger.Lazy) compiled clean and passed a dex-level bytecode check, but the .get() call still ran synchronously inside the constructor, so the cycle re-entered at runtime as a genuine StackOverflowError that crash-looped SystemUI on the owner's actual physical device after flashing — caught only because a real boot test was insisted on despite the build and bytecode check both passing; the corrected fix (f3f3c9ebc980) defers the call into a cold flow{} builder so it only runs after Dagger's DoubleCheck has memoized the singleton.

Signed with the alpha key (fingerprint 265ade399a0f3ee65003e56daef26a8f4d70e6a4), 13/13 AVB checks independently verified on two machines with two different avbtool binaries.

Task 8's real hardware evidence (docs/devices/lynx-duress-decoy-prelock-audit.md's own "Task 8" section, pasted here, not summarized): flash verification PASS (fingerprint, verified boot state, zero crash/ANR/tombstone over 31+ minutes); ADR-061 gate PASS (LockSettingsService: Verifying primary lockscreen credential for user 11 → Successfully verified, no device-wide bypass); Component A PASS 4/4 surfaces (lock screen, bouncer, QS, power menu — zero multi_user/avatar matches); owner-regression PASS (multi_user_switch icon enabled=true immediately after a lock→unlock with zero Settings.Global writes, Select-user dialog shows both "Owner" and the decoy — this is Task 6's fix, directly validated on hardware, closing Task 13's own prior narrower residual that the dialog's content itself had never been exercised); Component B real scenario PASS (one credential entry, no second prompt, real user stayed RUNNING_LOCKED, decoy reached RUNNING_UNLOCKED); Component B failure path UNCHECKED (source-read CachedAppOptimizer.forceFreezeAppAsyncLSP — fundamentally an async handler-message post with no organic synchronous RuntimeException path; Task 5's fix is real, correct plumbing, but the catch block itself has no safe way to trigger under real conditions); Component C latency PASS (warm n=5, 727-735ms, median 732ms; cold n=3 not 5, 1311-1350ms, median 1322ms, two further attempts lost to disclosed uiautomator timing flakiness, not a feature defect); all seven ADR-082 avatar-color patches confirmed present on the branch, the two never-before-exercised ones (0037, Multiuser 0003) confirmed at the bytecode level (UserManager.isMainForUiDisplay(I)Z) in the real built Settings.apk/Multiuser.apk, live pixel-color observation UNCHECKED (tooling limitation — uiautomator has no color data, screencap already documented unreliable for this Compose SystemUI).

Squashed-patch record (Task 1's audit, docs/devices/lynx-duress-unlock-patch-audit.md): three dead-end sequences reduced frameworks-base+build-release's combined patch count from 104 to 102 net (before Tasks 5-9's own additions) — build-release's credential_capture enable/revert cycle (four patches → one, 0ded9b0e3972), a frameworks-base resource flip reverted inside a later, differently-purposed patch (two patches' worth → dropped, real squash commit 64187dc103b5, correcting ADR-091 as above), and a pure path-fix follow-up (two patches → one, babea734ee48).

Accepted residual risks, each already established and re-affirmed here, not newly discovered: (1) the timing side-channel work was dropped (ADR-088) — the owner's own hard constraint (never slow the ordinary unlock) removed the only real mitigation; (2) Finding 13's re-lock sub-scenario (ADR-090) is proven untestable on this device — a stock GrapheneOS/AOSP Communal Hub gap, not an AegisOS defect; (3) 8 of 10 remaining pre-unlock concealment findings (Findings 9-12, 14-17) are source-verified code-absence facts with no dynamic test possible or needed; (4) ADR-072's six UNCHECKED leak-audit items have no trigger fixture on this device and match this project's own established precedent for that class of gap.

3Consequences

frameworks-base real tip: f3f3c9ebc980 (90 patches: 87 from ADR-091 + the three Task-7 patches above — 5c837bb3c965 the Kotlin migration, e6f6b9363b83 and f3f3c9ebc980 the two-attempt Dagger-cycle fix).

This is the first time in this plan's history that the candidate actually flashed, booted, and survived real-hardware verification — and the first attempt at that exact step (Task 7's own dex-checked build) silently shipped a crash loop despite every automated check passing, which is the strongest evidence yet in this project's own history for its standing rule that a build/bytecode check is not hardware verification.

RealUserQuiescerTest/UserRepositoryImplTest remain never executed via real atest — carried forward as a genuinely open item, not closed by assertion; the underlying diffs are small, direct, and independently source-verified, and Component B's real scenario (which exercises RealUserQuiescer.quiesce()'s success path directly) and the owner-regression evidence (which exercises UserRepositoryImpl's fixed flow directly) both provide real, if partial, dynamic coverage of the same code without the named test classes themselves ever running.

4Review when

Re-open this plan only for: (1) actually executing RealUserQuiescerTest/UserRepositoryImplTest via real atest, which needs either a build-server-reachable device or a working build-server emulator with adb attached — neither existed as of this closure; (2) Component B's failure path, if a safe, non-source-modifying way to trigger a real process-freeze failure is ever found; (3) live avatar-color pixel verification, if a reliable screenshot method for this Compose SystemUI is ever established; (4) either open UX gap this plan never touched (the owner→decoy swipe residual, the decoy→owner double-password path) — both explicitly out of this plan's scope per docs/devices/lynx-duress-decoy-prelock-audit.md's own "Task 15 (partial)" section, not silently forgotten.

Read from docs/governance/decision-log.md, entry ADR-092.