- Decision
- ADR-088
- Date
- 2026-09-17
- Status
- DroppedDropped, not implemented — the design document is kept as a record of what was investigated, not as a pending plan
- Part
- Duress Unlock
- Mentions
- ADR-087
- Mentioned by
- ADR-092, ADR-217, ADR-222, ADR-228, ADR-230
1Decision
Drop the timing-side-channel work docs/superpowers/specs/2026-09-17-duress-decoy-timing-overlap-design.md designed, after the owner's own hard constraint (never slow the ordinary unlock) and real evidence together removed the basis for continuing it now.
2Why and evidence
ADR-087 recorded the timing side channel as a single-trial measurement (771ms ordinary vs "~2050ms" duress, itself never arithmetically reconciling with its own stated 438+493+1110+809=2850ms breakdown).
This session re-measured the identical candidate (fade68916b06) directly on physical lynx across two reboots and found ordinary unlock stable (785-893ms) but duress unlock varying 1284-2764ms run to run — a spread larger than any plausible fix, including one unexplained 1347ms gap between the decoy launcher's process start and the actual user switch that nothing in the current architecture explains.
Separately, and independently disqualifying, read-only research against the live tree found the design's proposed mechanism (pre-start the decoy home task via onTop=false during the switch instead of after it) unsafe as scoped: the only cross-user visibility gate in this tree, ActivityRecord#showToCurrentUser() -> UserVisibilityMediator (mCurrentUserId == userId), is already satisfied before a pre-start could occur, because mCurrentUserId flips to the decoy's id as soon as the decoy user process starts (UserVisibilityMediator#assignUserToDisplayOnStart) — necessarily earlier than switch-complete.
onTop=false only sets launch z-order; it provides no compositing protection during that window. A safe version needs new, security-critical WindowManager logic (a per-task suppress-compositing flag, cleared only on confirmed switch-complete) that is its own design and its own security review, not an addition to this one.
Separately confirmed: TO_GONE_DURATION (500ms/200ms, FromPrimaryBouncerTransitionInteractor.kt) is a fixed-duration animation not gated on launcher draw-readiness, so even the design's own safe fallback (process-pre-fork-only, no early window) would not shorten the visible transition — only avoid a possible blank-frame glitch during it.
3Consequences
The timing side channel remains open and unmitigated, exactly as ADR-087 left it — this entry does not close it, it records that this session's specific attempt to partially close it is abandoned. The credential-check and grant-authorization mechanics investigated for this work are sound and reusable (DecoyProfileHelper.performDecoySwitch, UserController's grant record and intervening-switch invalidation at lines 2775/2810/2932/2973) should a future session pursue the larger WM-suppression redesign instead. No code changed; nothing was built, flashed, or landed.
4Review when
Revisit only if a future session is willing to scope a dedicated WindowManager-suppression design (new per-task compositing gate) with its own security review, or if the owner reconsiders the constraint that the ordinary unlock may never be slowed — either would change the calculus recorded here. Do not re-attempt the onTop=false pre-start mechanism as scoped in the dropped design without first addressing the showToCurrentUser() timing gap directly.
Read from docs/governance/decision-log.md, entry ADR-088.