- Decision
- ADR-059
- Date
- 2026-08-28
- Status
- AcceptedAccepted (qualified — two real open items, see Consequences)
- Part
- Duress Unlock
- Mentions
- ADR-047, ADR-049, ADR-058
- Mentioned by
- ADR-062, ADR-063, ADR-064
1Decision
Complete Tasks 8-9 of docs/superpowers/plans/2026-08-21-duress-decoy-user-switch.md — physical Pixel 7a hardware validation, real PIN-to-landing latency and battery measurement, and findings recording — closing out all 9 tasks of the original decoy-switch plan.
2Why and evidence
Real hardware validation on lynx (build path: frameworks/base HEAD 295f6f9cf610 → instrumented 49e4d939dc61 → instrumentation removed 5f3a562a4eef, each build signed with the alpha key over the full coherent 12-partition set per ADR-049's established method, 13/13 AVB-verified on two independent machines each time) confirmed with direct device evidence: real match, decoy match (via both cmd lock_settings verify and, incidentally, genuine physical keyguard PIN entry witnessed in logcat and confirmed by the owner as their own real interaction), real-user freeze (44/57 u0a* processes isFrozen=true), wrong-PIN handling (both real and decoy checks confirmed running via logcat), and the return-to-real-user path (ReturnToRealUserHelper, its first-ever physical-hardware exercise) all worked as designed.
Real PIN-to-landing latency measured directly via temporary IUserSwitchObserver instrumentation (UserSwitchObserver subclass, verified against real AOSP source before use so it could not itself stall the real switch): three same-boot runs, 1549ms/1120ms/643ms, average ≈1.1s with a clear cold/warm split — no target threshold exists in the design to score against, but this does not read as instantaneous, especially cold.
Real battery draw was measured via a detached on-device sampling script (necessary because the device's WiFi radio was found to fully disconnect during unplugged idle, and a first sampling attempt was discovered contaminated — still plugged in — and discarded before being used); the valid sample showed quiesced-state draw dropping roughly 5-25x below the non-quiesced baseline (-123,897 uA) once settled, directionally supporting the "really important" battery-savings goal from ADR-058's design but too small and irregular a sample (10 points, gaps up to 93s from the device's own deep-idle CPU suspend) to state as a precise ratio.
Along the way, a process-wide gap recurred and was fixed for real: OFFICIAL_BUILD=true was omitted from the first Task 8 build despite the ADR-047 lesson already being written into AGENTS.md — caught only because the owner asked before flash commands were sent, not by any check performed here; AGENTS.md and the corresponding memory record were both rewritten from a conditional judgment call into an unconditional, mechanical two-step requirement (export the flag on every real-device build; grep the fresh installed-files.txt/installed-files-product.txt for all four gated apps before any signing ceremony, every time).
3Consequences
Task 8's own two temporary-instrumentation patches (add, then remove) were fully reverted and are not exported to aegisos-platform — the real branch's frameworks/base tree is byte-identical to before Task 8 for the touched file, so nothing new ships.
Two real, confirmed-but-unresolved findings are carried forward, not fixed here: (1) the device's own multi-user UI chrome shows a visible two-profile indicator (distinct colors), a real-hardware confirmation of the "post-switch hiding" gap this project's own docs/research/duress-unlock-api-verification.md already flagged as out of scope for this plan — now known to be externally observable, not theoretical; (2) the decoy user does not land on a normal home screen after switching, most likely because it has never completed Android's own first-use setup flow (unconfirmed hypothesis).
The wipe-PIN path was not re-exercised on this physical device this session (owner declined a real, irreversible crypto-erase on an in-use test device) — its physical-hardware validation still rests solely on Stage 1 Task 3 Step 4 (2026-08-20).
4Review when
Revisit before treating Duress Unlock Stage 2 as feature-complete: both open UI findings need their own follow-up investigation and fix, and the battery number needs a longer, uninterrupted remeasurement before it can be quoted as a specification-grade result. Also revisit if a future real-device build ever again omits OFFICIAL_BUILD=true despite the now-unconditional AGENTS.md rule — that would mean the rule itself needs a stronger enforcement mechanism than documentation.
Read from docs/governance/decision-log.md, entry ADR-059.