Folding three Android apps into one Gradle build and one version list left the release APKs identical in size, manifest and dex SHA-1 before and after the build-logic step.
We folded three Android apps into one Gradle build, and the release files came out with the same size, the same manifest and the same dex SHA-1 as before. Dex is the bytecode file inside an Android app. SHA-1 is a fingerprint: equal fingerprints mean equal bytes. So the build restructuring left the shipped bytes where they were.
That claim covers one step. The network code that ABCam and ABSend share changed later, and the byte comparison does not cover it.
What we merged
Three apps, each kept with its full history: ABCam with 270 commits, ABDoc with 19 and ABSend with 29. A subtree copy brings another repository's history into a folder, so the history stays readable after the move.
One version list for the whole build. Kotlin, the language version, is 2.3.21 for all three. The template we started from had 2.2.21, but ABSend had already shipped on 2.3.21, so the list moved ahead of the template.
43 instruction files for the coding assistant, folded into one hub. We repointed 61 links to the new locations and found 0 dead ones.
A shared network module, once in one place, used by ABCam and ABSend. Both had been forks of the same hotspot and network tools.
The check that made it a claim
The release check ran at the build-logic step. We built the release APKs before the step and after it. Size, manifest and dex SHA-1 were identical. We signed with a throwaway key and checked the signature with apksigner, the Android signing checker.
This is a comparison of bytes, not a statement that the apps work. It says the build shape did not change what ships.
What the device checks found
The device checks ran the same day on a foldable test phone, and the second phone was used for a transfer test.
ABCam started its hotspot in 3,120 ms, and ABSend showed its hotspot as ready. The photo server answered with 200.
A second phone joined the ABSend hotspot and uploaded a 256 KB random file. The copy on the first phone had the same SHA-1 as the source.
An earlier line in our log said the ABSend hotspot had not run on a device. The device check later the same day ran it. We state the later result.
What failed, and what the byte check could not see
The screen recording found a bug the build checks did not. ABCam's in-app Start never moved the webcam service into the foreground. Only the adb path did that. So the app had no notification and no protection from the system stopping it. The fix is in commit bc0dee2. The draft in the Play console predates it, and its build number has to go from 6 to 7 before upload.
The order of the steps
The merge went in four steps, and each one had a way back.
The three repositories moved into one folder. Back-links stayed at the old paths, so older references still resolved, and a dock service that launches ABCam was repointed to the new path.
Each app's history came in through the subtree copy. The old repositories stayed on disk in hidden folders named for the pre-merge state, for rollback.
The build-logic step gave the three apps one shared build plugin for the SDK level, Java and Kotlin 17, and release signing and shrinking settings. This is the step with the byte comparison.
The network step moved the shared hotspot and network tools into one module, used by ABCam and ABSend.
What we still do not know
Whether the byte match holds after the network-module step. We did not repeat the release comparison there.
Whether the Flip's screen, which still read "Waiting for files" after the upload landed, does the same on the old build. We did not check.
Whether the ABCam desktop build works after the move. Its build script watches a git path under the old layout, and that path no longer exists. Nothing in our record says it failed, and nothing says it was tested.
How to prove this wrong
Build the release APK from the pre-merge repositories and from the merged one. Compare size, manifest and dex SHA-1. Then run the same comparison after the network-module step. If the bytes differ at that step, the claim covers only the build-logic step, and the post should say so.
Run the transfer test on a third phone with a different Android version.
Provenance
~/Brain/AB-Android/2026-11-17-consolidation.md: the subtree merge, the 43 instruction files, the 61 repointed links, the release comparison, the device and two-phone checks, and the foreground bug.
~/Projects/hardware/AB-Android/gradle/libs.versions.toml: the one version list and the Kotlin version note on lines 1 to 3.
Twelve network ideas tied, lost or starved against plain search or their control, with the counts for each. Five had a prereg file. The failures in our measuring that we caught are listed beside them.
On 88 frozen questions the cross-encoder reranker changed three held-out answers, all for the better, and lost none. A sign test on 3 to 0 gives p = 0.25, so the held-out gain is not yet established.
Commentaires
Pas encore de commentaires.
Se connecter pour commenter.