formatOf matched .tar.gz/.tgz as a combined format and then fell through to the bare .bz2 and .xz wrappers, so a .tar.bz2 was read as one compressed file: extracting it decompressed a single layer and left a .tar to extract again, while the same archive named .tar.gz unpacked in one step. The store listing advertised "extract ZIP, TAR, TAR.GZ, GZ, BZ2, XZ and 7z", which reads as though they all behave alike.
Found by extracting one archive of every supported format through the UI on the API 35 AVD rather than trusting the suite. A real z4.tar.bz2 left z4/z4.tar - raw ustar headers and all - beside z3.tar.gz fully unpacked to z3/a.txt and z3/sub/b.txt.
TAR_BZIP2 and TAR_XZ now sit alongside TAR_GZIP, detection consumes the whole tar suffix (.tar.bz2/.tbz2, .tar.xz/.txz) before the bare compressor, and the engine gains the two readArchive branches. The bare spellings still mean a single compressed file.
The suite had a gap that let this hide: only ZIP, GZIP, XZ and 7z had round-trip tests, so neither plain TAR nor TAR_GZIP did, and there was nothing for the asymmetry to contradict. All four tar spellings have one now, built in memory with the same library that reads them. 212 JVM tests, lint clean.
Also records the format sweep in EDGE_CASE_TESTING and re-measures the release artefacts (aab 4 838 235 / sha256 e1d23130..., apk 2 155 455 / sha256 37fb46cb...). The APK byte count happened to land on the same figure as the previous build, so the new literals were confirmed present in the shipped dex rather than trusted to the size.
The phone runs Android 12, so the questions that matter most for a release were exactly the ones it could not answer. Installed an API 35 AVD (sdkmanager was already present and dl.google.com reachable) rather than waiting for a public test track to report back. It found two user-visible defects, neither of which any test could have caught.
On Android 13+ the app asked for no runtime permission at all, so the progress notification was dead. StorageAccessPolicy.missingPermissions returned an empty list from API 33 on the reasoning that "all files access" is a special permission and browsing therefore needs no runtime grant — true as far as browsing goes, but the early return also dropped POST_NOTIFICATIONS, and nothing else in the app requested it. startForeground still ran, so nothing crashed and all 203 tests passed, while dumpsys showed granted=false and importance=NONE. Settings could not fix that either, because Android will not let a user grant a permission the app has never asked for. The notification grant is now requested alongside the storage grant, which is the only moment Android accepts it.
The four READ_MEDIA_* declarations were in the same position, and the manifest comment claiming they back thumbnails and viewers on Android 13+ was wrong. Coil loads both from a real filesystem path — the code comments already say "all-files mode only" — and the one MediaStore use is a best-effort delete request that "all files access" already covers. They were never requested, so they were never granted, so nothing could have depended on them. Removed, which also settles the checklist rule that declaring an unused permission is itself a policy violation.
The selection bar drew its last three actions off the screen on a phone. With one archive selected it holds ten icon buttons, ~480 dp at the 48 dp minimum touch size against the 411 dp a phone has, and BottomAppBar is a plain Row that does not scroll. Compose laid the surplus out past the right edge: measured on a 1080 px screen the row ran Copy(167) to Compress(923-980, itself squeezed to 57 px) and stopped, so Extract here, Select all and Clear selection were unreachable, while the same selection in a 2400 px landscape window showed all ten. The logic was never wrong — ArchiveFormats.isExtractable had judged the .zip correctly — the app just looked like it could not extract. The action row now scrolls, verified by swiping it in portrait and seeing all ten appear.
Verified on the AVD rather than only built: the notification dialog now appears and importance goes NONE -> DEFAULT, a real 220 MB copy posts notification 1001 on channel file-operations with ONGOING_EVENT|FOREGROUND_SERVICE and tears it down plus stops the service when done, edge-to-edge is correct under API 35s enforcement, predictive back returns from a folder and keeps the activity resumed, and Compress to ZIP produced an archive the host unzip read back with the right relative paths and contents. 204 JVM tests, lint clean, release artefacts re-measured.
The staged-delete work left a hole. The undo handle lived in BrowserUiState, so a delete from an earlier session was unreachable even though the files were sitting on the volume; and because nothing on disk recorded which folder an item came from, not even the user could put one back by hand with confidence. Adding a Trash screen alone would not have fixed it: the recovery data did not survive the process. So the origin is now written at delete time.
data/ gains its first persistence: a SQLite index (staged path, original path, name, type, size, timestamp). SQLite rather than Room because an annotation processor is a large addition to this build for one table and three queries, and behind a TrashIndex port so domain/Trash keeps no Context while the tests drive it with a map and the real one is still tested against real SQLite. New Trash entry in the overflow menu lists batches and restores a whole one.
The index is an index, not the truth: the staged files are the truth, and the two can disagree if a user empties the folder by hand or pulls a card. Reconciled both ways - restorableFrom drops entries whose file has gone, purgeExpired forgets entries whose folder has gone and sweeps folders no entry mentions - and rows are written after the move, so a crash leaves an unindexed folder (harmless, swept) rather than an entry that could never restore. Verified on hardware: delete, force-stop, relaunch, restore.
Also corrects the docs where they claimed input swipe x y x y 900 is delivered as a plain tap on MIUI. It does open the selection bar - it was driven that way - but it has also landed as a tap, so it is flaky rather than impossible.
SoftwareApp