- Decision
- ADR-049
- Date
- 2026-08-25
- Status
- AcceptedAccepted
- Mentions
- ADR-048
- Mentioned by
- ADR-050, ADR-059, ADR-062, ADR-072
1Decision
Resolve the lynx otacerts/test-key incident (docs/devices/lynx-ota-otacerts-testkey-incident.md) by choosing ADR-048's option (b) — flash the run-4 image bundle's own coherent 12-partition set as-is, rather than option (a)'s per-package shared-UID enumeration. Owner confirmed this choice before execution. Real device flash confirms the fix.
2Why and evidence
Before flashing, the run-4 bundle (<build-root>/signed/2026082507/lynx-img-20260825165502.zip, ADR-048's real success) was inspected read-only, no vault involved: unzip -l confirmed it contains all 12 AVB-relevant partition images from one single sign_target_files_apks-processed build (plus unrelated OEM firmware blobs and super_empty.img, correctly excluded from this fix — those are AOSP build pass-throughs, not part of this project's signing domain).
Contrary to the caution recorded when sign-baseline-images.sh was designed ("not trusting img_from_target_files's own auto-generated vbmeta.img as-is, since its rollback-index handling is unverified"), direct inspection found the bundle's own vbmeta.img was already correctly signed: public key cd7b74b3fb5c84cfc518ac961cc066d6473a4722 and rollback index 1783209600 both matched the device's currently-flashed, already-trusted vbmeta-v3.img exactly (confirmed by running avbtool info_image on both, not assumed).
This means ADR-048's real flash failure was never an AVB or key problem — the earlier manual add_hashtree_footer re-sign of just system.img was unnecessary rework; the actual defect was solely the 11-old-plus-1-new partition mix.
The full 13-file set (vbmeta.img + 12 partition images) was independently verified 13/13 on two machines (the build host's own avbtool, and a second copy on this workstation after an sha256-confirmed byte-identical scp transfer) before the owner flashed it via fastboot (all commands run by the owner directly, per this project's standing practice, never via an agent's Bash tool).
Real result, from the owner's own terminal output: all 13 fastboot flash operations OKAY, clean reboot, ro.build.version.incremental=2026082507 (the device now reflects vendor's real state too, unlike prior partial-flash sessions where only system changed), 0 Dropbox entries, and adb pull+openssl x509 on the live device's own /system/etc/security/otacerts.zip shows subject=CN=GrapheneOS — the AOSP test cert is gone from the device's real trusted store.
3Consequences
The otacerts/test-key security gap (device trusting anyone with a copy of AOSP source to sign an OTA) is closed on the physical device, confirmed by direct on-device evidence, not inferred from a clean flash exit code. sign-baseline-images.sh itself needed no changes for this — the fix was entirely in what got flashed (all 12 coherent partitions, not a splice), not in how the bundle was produced.
A durable finding for docs/research/build-and-signing-engineering-notes.md: img_from_target_files output, when driven through a properly --avb_*-keyed release-signing ceremony, can already carry a correctly-signed vbmeta.img with the right rollback index — the safe default is still to verify this directly (avbtool info_image/verify_image) rather than assume it, but it should not be assumed wrong either.
Task 9 of the OTA update server plan is not yet fully complete: this flash was applied via direct fastboot, not via the Updater app, so the already-published OTA (build 20260825082317) is now a no-op from the device's new position — the real end-to-end, app-driven OTA install (Task 9 Steps 5-6's actual purpose) still needs a fresh, newer candidate built and published now that the trust root is fixed.
4Review when
Revisit if a future candidate's img_from_target_files output does not self-verify cleanly (different device, different release-signing invocation) — this finding is about what was observed for lynx/this signing pipeline, not a general guarantee. Before Task 9 is marked complete: build, sign, and publish one more candidate, and have the Updater app perform a real install (not a direct fastboot flash) to confirm RecoverySystem.verifyPackage() now accepts the real release key in practice, not just in isolated static verification.
Read from docs/governance/decision-log.md, entry ADR-049.