- Decision
- ADR-062
- Date
- 2026-08-30
- Status
- AcceptedAccepted, partial — build and signing complete and independently verified; flashing, on-device validation, and closing findings (Task 11 Steps 2-4) not yet done as of this entry
- Mentions
- ADR-049, ADR-059, ADR-061
- Mentioned by
- ADR-060, ADR-063, ADR-072, ADR-079
1Decision
Build and sign a physical lynx hardware-validation candidate for this plan, rescoped per ADR-061's revert to validate only Components B, C, and D — Component A is intentionally not re-tested on physical hardware since it is reverted to inert. Flashing is handed to the owner to execute personally, per this project's standing rule that an agent never runs fastboot flash.
2Why and evidence
Built from all four repositories' current real state (frameworks/base 3941aa4131fe, build/release 10209de65d2accda1d7cdb8e178f41f9a20f166f with Component A's flag DISABLED, packages/apps/Settings 229434697ac, packages/apps/Multiuser 0ab4e93) via lunch lynx-cur-userdebug, export OFFICIAL_BUILD=true in the same command block as OUT_DIR/envsetup.sh — an incremental build against the warm out-lynx tree, chosen deliberately over a fresh OUT_DIR given only 49GiB free on /srv against this project's own ~92GB cold-build-cost precedent; disk stayed at 49-50GiB free throughout, confirming the choice.
OFFICIAL_BUILD=true's effect was independently confirmed by the controller directly on the build server, not just accepted from the implementer's report: Updater/Auditor/Seedvault/PdfViewerGOS all present in the fresh installed-files*.txt, with a build-matching mtime — grepping the build log for the literal string is not a meaningful test, since OFFICIAL_BUILD is a Makefile conditional never echoed anywhere.
The signing key was determined from real evidence rather than assumed: the device's actual current AVB trust root was confirmed, via avbtool info_image against the most recent real signing evidence anywhere on the server (duress-task8-sign-workdir/vbmeta.img), to be the original alpha key (265ade399a0f3ee65003e56daef26a8f4d70e6a4) — not the signing host's vault key that ADR-049/050 had moved the separate OTA-publish path to — resolving an apparent contradiction in ADR-059's own prose in the implementer's favor (it was read literally rather than assumed to be loose terminology, and turned out correct).
All 12 partition images plus a freshly signed vbmeta.img (rollback index 1783209600, unchanged) were independently verified 13/13 on two separate machines (the build host's AOSP-tree avbtool.py and this workstation's distinct /usr/bin/avbtool install); the controller independently re-ran this verification a third time, directly, against the locally staged files, before handing anything to the owner.
3Consequences
The full fastboot flash command sequence (14 commands, fastbootd for the dynamic partitions, vbmeta last, mirroring ADR-049's proven-correct ordering) was handed to the owner to run themselves; not executed by any agent. The verified image set's only durable copy is on the build host (<build-root>/evidence/task11-phaseA-sign-workdir/) — the locally staged copy used for flashing lives in this session's own ephemeral tmpfs scratchpad and will not survive a workstation reboot.
4Review when
Once flashed, Task 11 Steps 2-4 remain to be done: repeat Task 10's Component B/C/D checks with real evidence on physical hardware (explicitly skipping Component A's now-known-broken/reverted mechanism), decide whether the double-PIN latency remeasurement (the plan's original Step 3) is even meaningful now that Component A is inert, and write the closing findings record this plan's own eventual completion ADR will need. This workstation's /home being at 100% (1.5GiB free) was flagged as a real, current friction point for staging future signed images — worth addressing before the next hardware-validation cycle.
Read from docs/governance/decision-log.md, entry ADR-062.