Make the gate name the levels it ran instead of claiming all of them

The closing line was `green at every supported API level`, printed on both
paths -- including the one that had just said `NOT COVERED LOCALLY: API 37` two
lines above. A false claim, printed by the tool whose entire purpose is to stop
false claims reaching CI, on its first run.

It now names them: `green on API 33, 34, 35, 36` when the Pixel is absent, and
`green on API 33, 34, 35, 36, 37` when it is attached and passed.

Nothing else changes. The app/src subtree is untouched, so this exercises the
cache scoping from the previous commit: the sweep is skipped as already green and
only the JVM gate runs -- which is the whole reason that key was moved off the
repo tree.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.github/workflows/status_check.yml

@@ -272,13 +272,15 @@ jobs:
           # has to begin after them, not between them. The name is stale and kept:
           # read .github/scripts/e2e-run.sh's header, which carries the measurements.
           #
-          # notAnnotation below keeps five tests off this row, and one of
-          # them is new. SafPickerRoundTripTest's PICKER test was measured on
-          # 2026-08-24 as passing here and was left on the leg; four gating logcats
-          # read on 2026-09-05 show it aborting system_server from the task-snapshot
-          # path on every single run, pass or fail, which is what had been failing
-          # unrelated PRs (#108). Both of that class's tests now carry the marker.
-          # docs/api-37-emulator-crash.md has the timings and the correction.
+          # notAnnotation below keeps seven tests off this row. SafPickerRoundTripTest's
+          # PICKER test was measured on 2026-08-24 as passing here and was left on the
+          # leg; four gating logcats read on 2026-09-05 show it aborting system_server
+          # from the task-snapshot path on every single run, pass or fail, which is what
+          # had been failing unrelated PRs (#108). All FOUR of that class's tests now
+          # carry the marker -- the two saves through the picker (#226, #250) joined on
+          # 2026-09-06 by inheritance rather than measurement, since they open the same picker.
+          # docs/api-37-emulator-crash.md has the timings and the correction, and
+          # FailsOnEmulatorApi37.kt has why the third one cannot be measured here.
           #
           # api-level must be a POINT release. A bare 37 is not an SDK package and
           # fails during setup, which cost a run to discover. `37.0` is the choice
@@ -288,9 +290,12 @@ jobs:
           # docs/api-37-emulator-crash.md measures 37.0 rev 6 and 37.1 rev 8 side
           # by side, so pinning 37.0 is a decision, not a constraint.
           #
-          # notAnnotation removes the three tests that do not pass on this image; they
+          # notAnnotation removes the seven tests that cannot be RUN on this image; they
           # run in the advisory job below, off the same marker so they cannot end up
-          # in both or neither. docs/api-37-emulator-crash.md has the measurements.
+          # in both or neither. "Cannot be run" rather than "do not pass" is deliberate:
+          # four fail outright, one of those aborts the framework on its way down, and on
+          # the advisory leg the three picker tests behind it never report at all.
+          # docs/api-37-emulator-crash.md has the measurements.
           - label: "37"
             api-level: "37.0"
             disable-system-ui: "1"

CLAUDE.md

@@ -76,12 +76,21 @@ days. Read it as the current answer, and see the git history if you need the old
   `angle_indirect` and `swangle_indirect` all boot, while `auto`, `off`, `guest` and
   `swiftshader_indirect` do not. `docs/local-emulator.md` has the evidence and the per-API renderer
   table.
-- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. **Five** of the 69 instrumented
-  tests cannot be *run* on that image, for three unrelated reasons: three Media3 tests fail inside
-  the emulator's own `c2.goldfish.h264.decoder`, one SAF test takes the framework down when it
-  rotates the display, and its sibling — the SAF picker round trip — aborts `system_server` from
-  the task-snapshot path whether it passes or not. All five carry `@FailsOnEmulatorApi37` and run
-  in a separate `continue-on-error` job; the gating leg runs the other 64.
+- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. **Seven** of the 71 instrumented
+  tests cannot be *run* on that image, for three measured reasons and two inherited: three Media3
+  tests fail inside the emulator's own `c2.goldfish.h264.decoder`, one SAF test takes the framework
+  down when it rotates the display, and its sibling — the SAF picker round trip — aborts
+  `system_server` from the task-snapshot path whether it passes or not. The sixth, that class's
+  two saves through the picker (#226 and #250), carry the marker because they open the same picker
+  and a second DocumentsUI dialog on top of it — **not** because either has ever been observed here. It cannot be:
+  the rotation test runs first and takes the framework down, so all five advisory runs at the
+  previous baseline reported `expected: 6, received: 4, failed: 4`, and the four were the three
+  Media3 tests plus the rotation — runs 34041156680, 34041593697, 34042397320, 34043502322 and
+  34045105857. **No picker test has ever reported on the advisory leg**, which is a correction to
+  what the marker's own KDoc used to say. All seven carry `@FailsOnEmulatorApi37` and run in a
+  separate `continue-on-error` job; the gating leg runs the other 64 — **the same 64 for the third
+  time running**, which is exactly how this paragraph goes stale unnoticed: 69−5, 70−6 and 71−7
+  are all 64.
 
   **These two numbers move with the suite and are derived, not remembered.** `grep -cE
   '^\s*@Test' ` over `app/src/androidTest` is the first; the second is that minus the marker
@@ -114,7 +123,7 @@ days. Read it as the current answer, and see the git history if you need the old
   describes everything in it. The name is kept deliberately — it is not a required context and
   people have learned to look for it — so **read the marker, not the name**, for what it holds.
   **It is red on every PR, by design**: do not read it as your change breaking something, and do
-  not read a green run as evidence those five tests pass.
+  not read a green run as evidence those seven tests pass.
   `docs/api-37-emulator-crash.md` has the measurements.
 
   **That instruction is also why nobody looks, so the job now reports its own shape** — expected,
@@ -134,7 +143,7 @@ days. Read it as the current answer, and see the git history if you need the old
   is gradle never returning, so the log it left says nothing about it.
 
 Still true, and the reason the advisory job is not simply deleted: **API 37 needs a manual check on
-the Pixel 10 Pro XL before each release.** Those five tests are the one thing CI cannot answer
+the Pixel 10 Pro XL before each release.** Those seven tests are the one thing CI cannot answer
 for.
 
 On a device or emulator, build only the ABI it can execute:
@@ -339,7 +348,7 @@ install for code that can never run — and on API 37 the full APK does not fit
   when a fix is for something intermittent.
 
   **Every number above is `testDebugUnitTest` only, and on 2026-09-05 the instrumented suite got its
-  first read for that reason** — `docs/e2e-read-findings.md`, entries **E1-E6**, tickets
+  first read for that reason** — `docs/e2e-read-findings.md`, entries **E1-E7**, tickets
   **#223-#230**. Four waves had been steered by a figure that **cannot see `app/src/androidTest` at
   all**, so nothing had ever asked what those 60 device tests pin, only that they were green.
 
@@ -378,6 +387,45 @@ install for code that can never run — and on API 37 the full APK does not fit
   half. But the *input* bridge needs no documents provider at all, which is what kept #225 headless
   and is how #238 surfaced.
 
+  **The 2026-09-06 re-check found that the read's own last PR had re-introduced the drift the read
+  was about**, and that is the entry worth carrying forward. #226 moved the suite 69 -> 70 and the
+  markers 5 -> 6 and changed neither the count in this file, the marker's KDoc, nor the two
+  comments in `status_check.yml`. **The gating figure is what hid it**: 69 - 5 and 70 - 6 are both
+  64, so the one number a reader checks against a run had not moved — which is precisely why the
+  paragraph above says to derive these rather than remember them. Worse, two KDoc claims in the new
+  test described a draft rather than the code: it says MP3 was chosen so the setup could not depend
+  on the device's codecs, while the code converts at the default `MP4_H265`/`FAST` and therefore
+  routes on `canEncode(H265)` — the *negation* of the stated reason. **That is E1 and E3's failure
+  mode, committed by the wave that found it.** All of it is fixed; the standing item is **#250**,
+  because #226 proved D4's premise and never drove its delete arm.
+
+- **Nothing is committed or pushed until the local gate is green, at every supported API level.**
+  Source work (`app/src/main`) needs the unit tests **and** the instrumented tests passing on every
+  level; test work (`app/src/test`, `app/src/androidTest`) needs the whole suite passing on every
+  level. `tools/git-hooks/local-gate.sh` enforces it as `pre-commit` and `pre-push`; wire it up once
+  with `git config core.hooksPath tools/git-hooks`.
+
+  33-36 run the whole suite on emulators. **API 37 cannot be run on an emulator on this host at
+  all** — not "is red", *cannot run*: measured 2026-09-06, the image logs `3 new surfaceflinger
+  aborts in 45 s (want 0)` and then the APK install itself fails with `Can't find service:
+  package`, because the framework is gone before Gradle installs anything. `Starting 0 tests`. So
+  the hook runs API 37 on the **attached Pixel 10 Pro XL** when it is there, and says plainly that
+  the level is uncovered when it is not — CI's gating leg being what answers for it then. It never
+  claims five levels having run four.
+
+  The sweep is cached under the hash of the **`app/src` subtree**, not the whole repo tree. Keying
+  it on the whole tree was the first cut and it was wrong: editing a comment in `CLAUDE.md` threw
+  away a sweep of byte-identical application code and re-ran forty minutes of emulators to prove
+  nothing, which is how a gate teaches people to resent it. Any change under `app/src` still
+  invalidates it, and the JVM gate runs unconditionally. **There is deliberately no skip
+  variable**, and `--no-verify` needs the repo owner's say-so each time rather than being reached
+  for when the gate is inconvenient.
+
+  Why it is worth tens of minutes a commit: the alternative was measured on 2026-09-06, when one PR
+  spent several gating legs learning one leg at a time what a sweep answers in one pass — and the
+  failing leg **moved** between runs (API 35 red then green, API 34 green then red). One leg at a
+  time that reads as someone else's flake; as a sweep it is one signal.
+
 - **Testable code is not done until it is tested.** If a piece is unit testable, it gets unit
   tests before it counts as done. If it is e2e testable, it gets e2e tests. Both clauses apply —
   a change that is both needs both.

app/src/androidTest/java/org/libremediaconverter/FailsOnEmulatorApi37.kt

@@ -9,14 +9,23 @@ package org.libremediaconverter
  * That is the whole reason there is one annotation rather than a pair of test lists: two lists
  * drift, and the drift is silent in both directions (a test that runs nowhere reads as green).
  *
- * **"Cannot be run" covers two things, and it said only the first until 2026-09-05.** Four of the
- * five carriers simply fail: three Media3 tests die in the image's own `c2.goldfish.h264.decoder`,
- * and the SAF rotation test takes the framework down with it. The fifth —
- * `SafPickerRoundTripTest.pickingAFileThroughTheSystemPickerFillsInTheFileCard` — **passes about
- * half the time and aborts `system_server` every time**, which is worse for a gating leg than an
- * honest failure: it fails the leg from the teardown, with no failing test to point at (#108).
- * The wording was widened rather than the test excused; that test's own KDoc has the four-run
- * measurement.
+ * **"Cannot be run" covers three things now, and it covered only the first until 2026-09-05.**
+ * Four of the seven carriers simply fail: three Media3 tests die in the image's own
+ * `c2.goldfish.h264.decoder`, and the SAF rotation test takes the framework down with it. The
+ * fifth — `SafPickerRoundTripTest.pickingAFileThroughTheSystemPickerFillsInTheFileCard` —
+ * **passes about half the time and aborts `system_server` every time**, which is worse for a
+ * gating leg than an honest failure: it fails the leg from the teardown, with no failing test to
+ * point at (#108). The wording was widened rather than the test excused; that test's own KDoc has
+ * the four-run measurement.
+ *
+ * **The sixth and seventh are the new third thing: they are marked by inheritance, not by
+ * measurement.** `SafPickerRoundTripTest.aSaveWritesToTheDocumentTheSystemPickerCreated` (#226)
+ * and `.aFailedSaveDeletesTheDocumentItCouldNotWrite` (#250) each open the same picker and then a
+ * second DocumentsUI dialog on top of it, so they sit on the same task-snapshot path their sibling
+ * was marked for. Neither has ever been observed at API 37 either way — see the measurement under
+ * [FAILS_ON_EMULATOR_API37_BASELINE], which is why they cannot be. Marking them was the
+ * conservative choice, and **the trigger for revisiting it is the rotation test, not themselves**:
+ * while that one truncates the advisory run, nothing downstream of it can report.
  *
  * It says only what has been measured: **on the emulator, at API 37.** The same tests pass on a
  * physical Pixel 10 Pro XL at API 37 and at API 33–36 on the same runner under the same renderer,
@@ -46,19 +55,21 @@ annotation class FailsOnEmulatorApi37
  * keep printing with nothing to compare to, so it announces that it could not read the baseline
  * rather than falling quiet. If you see that notice, this line is what it means.
  *
- * **One number, both checks, and that is what the marker means.** A test carrying it cannot be run
- * on this image, so the count is simultaneously how many the advisory leg runs and how many fail.
- * A *smaller* failure count is the interesting direction: it means one of them now passes, which
- * is the trigger the KDoc above names for deleting the annotation.
+ * **One number, both checks, and that is what the marker was meant to mean.** A test carrying it
+ * cannot be run on this image, so the count is meant to be simultaneously how many the advisory
+ * leg runs and how many fail. A *smaller* failure count is the interesting direction: it means one
+ * of them now passes, which is the trigger the KDoc above names for deleting the annotation.
+ * **Since 2026-09-06 the second half no longer holds in practice** — the run truncates before
+ * three of the seven start, which the last paragraph below measures. `expected` still holds, and it is the
+ * field that catches a marker added without changing this number.
  *
- * **The picker test is the one to read that sentence carefully for.**
- * `pickingAFileThroughTheSystemPickerFillsInTheFileCard` was marked on 2026-09-05 for aborting
- * `system_server` rather than for failing (#108), and on the gating leg it passed two runs of
- * four. It fails on the advisory leg because the rotation test runs before it and takes the
- * framework down first — measured, `api37-debug.yml` run 34008889182, which reports
- * `expected: 4, received: 4, failed: 4` with the four in the order Media3, Media3, rotation,
- * picker. (Those dispatches predate the third Media3 marker landing on `main`, so their totals
- * are four rather than five; the ordering they establish is what matters here.)
+ * **The picker tests are the ones to read that sentence carefully for, and the reason changed
+ * on 2026-09-06.** `pickingAFileThroughTheSystemPickerFillsInTheFileCard` was marked on
+ * 2026-09-05 for aborting `system_server` rather than for failing (#108), and on the gating leg
+ * it passed two runs of four. It was recorded here as *failing* on the advisory leg, behind the
+ * rotation test — measured, `api37-debug.yml` run 34008889182, `expected: 4, received: 4,
+ * failed: 4`, in the order Media3, Media3, rotation, picker. (Those dispatches predate the third
+ * Media3 marker, so their totals are four rather than six.)
  *
  * **But a second dispatch of the identical configuration reported 4/3/3**, having lost the last
  * test to the abort rather than to anything about the test list, and that is why
@@ -68,6 +79,15 @@ annotation class FailsOnEmulatorApi37
  * reporting fewer failures than this as one of them now passing; read a truncated one as the
  * framework having died, which is this job's normal.
  *
+ * **That is no longer what happens, and the difference is that neither picker test reports at
+ * all.** The rotation test truncates the run before them: **all five** advisory runs at the
+ * previous baseline of six — 34041156680, 34041593697, 34042397320, 34043502322 and 34045105857 —
+ * report `expected: 6, received: 4, failed: 4`, and the four are the three Media3 tests plus the
+ * rotation. #250 adds a third picker test behind the same wall, so expect `expected: 7,
+ * received: 4`. So the advisory leg currently answers for
+ * four of its six, and the comparison below is unaffected only because `failed` is not compared
+ * on a truncated run. Read it as **unmeasured**, not as passing or failing.
+ *
  * So: adding or removing a [FailsOnEmulatorApi37] means changing this number, in this file, in
  * the same diff. The report says so on the run itself if you forget — it prints the tree's own
  * `grep` count beside this one.
@@ -78,4 +98,4 @@ annotation class FailsOnEmulatorApi37
  * `INSTRUMENTATION_ABORTED`, so the count is a number taken from a partial run. The report
  * records the truncation next to the counts for that reason.
  */
-const val FAILS_ON_EMULATOR_API37_BASELINE = 6
+const val FAILS_ON_EMULATOR_API37_BASELINE = 7

app/src/androidTest/java/org/libremediaconverter/saf/FixtureDocumentsProvider.java

@@ -14,8 +14,6 @@ import java.io.FileOutputStream;
 import java.io.IOException;
 import java.io.InputStream;
 import java.io.OutputStream;
-import java.util.ArrayList;
-import java.util.List;
 
 /**
  * One file, offered to the system file picker, so that picking one can be tested at all.
@@ -116,7 +114,6 @@ public final class FixtureDocumentsProvider extends DocumentsProvider {
     public static final String DESTINATION_PREFIX = "dest/";
 
     /** Document ids {@link #deleteDocument} was called with, newest last. Cleared by {@link #reset}. */
-    private static final List<String> DELETED = new ArrayList<>();
 
     /** Already in this source set, and already a real H.264 MP4 the engines can open. */
     private static final String FIXTURE_ASSET = "sample_h264.mp4";
@@ -239,24 +236,11 @@ public final class FixtureDocumentsProvider extends DocumentsProvider {
         if (documentId == null || !documentId.startsWith(DESTINATION_PREFIX)) {
             throw new FileNotFoundException("refusing to delete: " + documentId);
         }
-        synchronized (DELETED) {
-            DELETED.add(documentId);
-        }
         destinationFile(documentId).delete();
     }
 
-    /** Document ids {@link #deleteDocument} was called with, newest last. */
-    public static List<String> deletedDocumentIds() {
-        synchronized (DELETED) {
-            return new ArrayList<>(DELETED);
-        }
-    }
-
-    /** Forgets recorded deletes and removes created destinations. The process outlives one class. */
+    /** Removes created destinations. The process outlives one class. */
     public static void reset(File filesDir) {
-        synchronized (DELETED) {
-            DELETED.clear();
-        }
         File dir = new File(filesDir, "destinations");
         File[] children = dir.listFiles();
         if (children != null) {

app/src/androidTest/java/org/libremediaconverter/saf/SafPickerRoundTripTest.kt

@@ -25,9 +25,11 @@ import androidx.test.uiautomator.Configurator
 import androidx.test.uiautomator.StaleObjectException
 import androidx.test.uiautomator.UiDevice
 import androidx.test.uiautomator.Until
+import androidx.work.WorkManager
 import org.junit.After
 import org.junit.Assert.assertArrayEquals
 import org.junit.Assert.assertEquals
+import org.junit.Assert.assertFalse
 import org.junit.Assert.assertNotEquals
 import org.junit.Assert.assertNotNull
 import org.junit.Assert.assertTrue
@@ -40,6 +42,7 @@ import org.libremediaconverter.convert.ConversionDependencies
 import org.libremediaconverter.convert.OutputPublisher
 import org.libremediaconverter.ui.TestTags
 import java.io.File
+import java.io.OutputStream
 import java.util.concurrent.atomic.AtomicInteger
 import java.util.regex.Pattern
 
@@ -280,17 +283,41 @@ private class RecordingPublisher(private val app: Context) : OutputPublisher(app
         super.publish(staged, destination)
     }
 
+    /**
+     * Refuses the write when [failOpen] is set, which is the forcing condition for #250.
+     *
+     * Returning null rather than throwing is deliberate: it is the arm `publish`'s
+     * `?: error("Could not open destination for writing")` exists for, and `openDestination`'s
+     * own KDoc says a provider that is present and declines is the half no fake can produce on
+     * demand. The size probe in `publish` has already run by the time this is reached, so
+     * `destinationWasEmpty` is true and `deletePartialOutput` is reached with the document
+     * genuinely empty — which is the whole point.
+     */
+    override fun openDestination(destination: Uri): OutputStream? =
+        if (failOpen) null else super.openDestination(destination)
+
     companion object {
         var savedBytes: ByteArray = ByteArray(0)
         var seenDestination: Uri? = null
         var seenIsDocumentUri: Boolean? = null
         var seenSizeBefore: Long? = null
 
+        /**
+         * Makes the next `publish` refuse to open its destination.
+         *
+         * A flag rather than a second publisher because `ConversionDependencies.publisher` is one
+         * seam and there is no orchestrator: every test in this process shares the instance the
+         * `init` block installed. [reset] clears it in teardown, so a test that sets it cannot
+         * leak a refusing publisher into the next class.
+         */
+        var failOpen: Boolean = false
+
         fun reset() {
             savedBytes = ByteArray(0)
             seenDestination = null
             seenIsDocumentUri = null
             seenSizeBefore = null
+            failOpen = false
         }
     }
 }
@@ -367,6 +394,8 @@ class SafPickerRoundTripTest {
     fun restoreOrientation() {
         // The suite runs without Android Test Orchestrator, so a swapped seam outlives the class.
         ConversionDependencies.reset()
+        RecordingPublisher.reset()
+        clearFinishedWork()
         ActivityLifecycleMonitorRegistry.getInstance().removeLifecycleCallback(recreationWatcher)
         if (!rotated) return
         device.setOrientationNatural()
@@ -515,9 +544,28 @@ class SafPickerRoundTripTest {
      *
      * ## The conversion is setup, not subject
      *
-     * Save is only offered on `Converted`, so the test converts first. MP3 is chosen because the
-     * router sends it to FFmpeg unconditionally at every API level, so the setup cannot depend on
-     * the device's codecs — #223 is what that costs.
+     * Save is only offered on `Converted`, so the test converts first, at the screen's default
+     * `MP4_H265` / `FAST`. That is **not** codec-independent, and this KDoc claimed the opposite
+     * until 2026-09-06: an earlier draft used MP3 for exactly that reason, and the format had to
+     * move for a different constraint the picker imposes — [convertToTheDefaultFormat] has it.
+     * `MP4_H265` at `FAST` reaches `ConversionRouter`'s `canEncode(H265)` gate, so it runs on
+     * FFmpeg on the emulators (no hardware H265) and on Media3 on the Pixel.
+     *
+     * **That is tolerable here, and #223 is the reason it needs saying.** There, the routing
+     * decided whether the *subject* was reached, so a route to FFmpeg made the test pass while
+     * proving nothing. Here the conversion is setup: if it goes the other way and fails, this test
+     * fails loudly on the setup rather than quietly on the assertion. The subject is what `publish`
+     * was handed, which the engine that produced the file does not touch.
+     *
+     * ## Why it carries [FailsOnEmulatorApi37]
+     *
+     * By inheritance, not measurement. It opens the same picker as
+     * [pickingAFileThroughTheSystemPickerFillsInTheFileCard], which was marked for aborting
+     * `system_server` from the task-snapshot path (#108), and then a second DocumentsUI dialog on
+     * top of it. It has never been observed at API 37 either way: the rotation test truncates the
+     * advisory run first, so all four advisory runs at this baseline report
+     * `expected: 6, received: 4` without reaching either picker test. Marking it was the conservative choice and it is
+     * recorded as unmeasured in `FailsOnEmulatorApi37.kt` rather than dressed up as a measurement.
      */
     @Test
     @FailsOnEmulatorApi37
@@ -551,6 +599,116 @@ class SafPickerRoundTripTest {
         assertTrue("staging should be empty after a successful save", staged.listFiles().isNullOrEmpty())
     }
 
+    /**
+     * The other half of D4 (#250): a save that fails deletes the document it could not write.
+     *
+     * ## Why this is separate from the test above
+     *
+     * #226 proved the *premise* — SAF hands back a document reporting exactly zero bytes, so
+     * `destinationIsKnownEmpty` can answer true — and then drove the success path, where the
+     * `catch` is never entered. So `deletePartialOutput` had still never run against a real
+     * `DocumentsProvider`; its only assertions were `OutputPublisherPublishTest`'s, against
+     * `FakeSafProvider` under Robolectric. That is the same "asserted only against a fake built to
+     * match it" shape #226 was filed to break, one layer down.
+     *
+     * ## The forcing condition, and why it is a returned null
+     *
+     * [RecordingPublisher.failOpen] makes `openDestination` return null. `publish` turns that into
+     * `error("Could not open destination for writing")` **after** its size probe has already run,
+     * so the `catch` is reached with `destinationWasEmpty == true` on a document DocumentsUI
+     * created seconds earlier. Nothing is simulated: the URI, the grant, the provider and the
+     * delete are all real.
+     *
+     * Null rather than a throw because `openDestination`'s KDoc says a provider that is present
+     * and declines is the half no fake can produce on demand — so this is also the first time that
+     * arm has been taken against a live provider rather than a stub.
+     *
+     * ## The oracle, and why it is not a recorder inside the provider
+     *
+     * The obvious assertion — have the provider record what `deleteDocument` was called with, and
+     * read it back — **cannot work here, and finding that out is half of what this test cost.**
+     * `FixtureDocumentsProvider` is declared by the test APK and runs in
+     * `org.libremediaconverter.test`; instrumentation runs in the app's process. A `static` in the
+     * provider is therefore a different object from the one a test can see, and the accessor #226
+     * left behind read empty on every run. That is E7's process wall from a third side, after
+     * `ACTION_OPEN_DOCUMENT` and `ActivityScenario`.
+     *
+     * So the oracle is the document, which does cross the boundary because the app holds a URI
+     * grant for it. **This is still the path rather than the artefact**, because the two
+     * assertions are read together: the size query above proves the document *existed and was
+     * empty* moments earlier, and a `content://` document that no longer answers a query is one
+     * something deleted. Nothing else in the app deletes SAF documents.
+     *
+     * The staged file is asserted to **survive**, which is the deliberate other half of that
+     * `catch`: a failed save may leave the staged copy as the only copy of an hour of transcoding,
+     * so `ConversionViewModel` keeps it and puts "Try saving again" on screen.
+     */
+    @Test
+    @FailsOnEmulatorApi37
+    fun aFailedSaveDeletesTheDocumentItCouldNotWrite() {
+        pickTheFixture()
+        convertToTheDefaultFormat()
+
+        RecordingPublisher.failOpen = true
+        saveThroughTheSystemPicker(settlesOn = TestTags.RETRY_SAVE)
+
+        val destination = RecordingPublisher.seenDestination
+        assertNotNull("publish was never reached, so the delete arm was not exercised", destination)
+        assertEquals(
+            "the document was not positively empty, so publish would refuse to delete it",
+            0L,
+            RecordingPublisher.seenSizeBefore,
+        )
+        assertFalse(
+            "publish did not delete the document it could not write: $destination",
+            documentStillExists(destination!!),
+        )
+
+        // The staged copy is kept on purpose -- see ConversionViewModel.save's onFailure.
+        val staged = File(context.cacheDir, "conversions")
+        assertTrue(
+            "a failed save must not delete the staged file; it may be the only copy",
+            staged.listFiles()?.isNotEmpty() == true,
+        )
+    }
+
+    /**
+     * Leaves nothing for the next test's launch to reattach to.
+     *
+     * **In teardown rather than at the end of a test, and that placement is the point.**
+     * `aFailedSaveDeletesTheDocumentItCouldNotWrite` proves that a failed save *keeps* its staged
+     * file — deliberately, since it may be the only copy — so it ends with a finished job and a
+     * live staged file, which is exactly what the app reattaches to on the next launch. Its
+     * sibling then opened on `Converted` with no "Choose file" to tap: measured, as a 30 s timeout
+     * on `converter.chooseFile` in a test that had nothing wrong with it.
+     *
+     * The first fix tapped "Start over" at the end of the test body. That works until the test
+     * fails, and then it does not run at all — measured too, on the mutation run that proved this
+     * suite bites: one real failure became two, and the second looked like an unrelated flake.
+     * **One cause must produce one red test**, so the cleanup belongs where it runs either way.
+     *
+     * **`pruneWork` and not `cancelAllWork`, on design grounds and not on a measurement.** Only
+     * finished work records need to go — that is all the next launch reattaches to — and
+     * `cancelAllWork` additionally cancels live work, which is a wider blast radius than teardown
+     * in a shared process needs. `pruneWork` cannot touch a job that has not run yet.
+     *
+     * `cancelAllWork` was **suspected** of causing an API 35 red here and did not cause it; see
+     * [CONVERSION_TIMEOUT_MS], which did. A local API 35 run with `cancelAllWork` passed, and the
+     * logcat showed the conversion encoding rather than cancelled. The narrower call is kept
+     * because it is the right one, not because it fixed anything.
+     */
+    private fun clearFinishedWork() {
+        WorkManager.getInstance(context).pruneWork()
+        File(context.cacheDir, "conversions").listFiles()?.forEach { it.delete() }
+    }
+
+    /** Whether [destination] still answers a metadata query. A deleted document does not. */
+    private fun documentStillExists(destination: Uri): Boolean = runCatching {
+        context.contentResolver
+            .query(destination, arrayOf(OpenableColumns.SIZE), null, null, null)
+            ?.use { it.moveToFirst() } ?: false
+    }.getOrDefault(false)
+
     /**
      * Runs the conversion, leaving the screen on `Converted`.
      *
@@ -617,7 +775,7 @@ class SafPickerRoundTripTest {
      * Retried whole, for the reason [pickTheFixture] documents: a dialog that came up unreadable
      * cannot be recovered from inside, and a fresh one is the only answer.
      */
-    private fun saveThroughTheSystemPicker() {
+    private fun saveThroughTheSystemPicker(settlesOn: String = TestTags.Converter.CONVERT_ANOTHER) {
         var missing: BySelector? = null
         repeat(PICK_ATTEMPTS) { attempt ->
             requireAReadableScreen()
@@ -626,7 +784,9 @@ class SafPickerRoundTripTest {
                 if (attempt == 0) PICKER_TIMEOUT_MS else REOPENED_TIMEOUT_MS,
             )
             if (missing == null) {
-                awaitNode(TestTags.Converter.CONVERT_ANOTHER, SAVE_TIMEOUT_MS)
+                // The node that says the save has *finished*, either way. Waiting on the success
+                // one when the save is meant to fail would time out on a test that is working.
+                awaitNode(settlesOn, SAVE_TIMEOUT_MS)
                 return
             }
             dismissThePicker()
@@ -1038,9 +1198,35 @@ class SafPickerRoundTripTest {
         }
     }
 
+    /**
+     * Waits for [tag], treating "the app has no composition right now" as *not yet* rather than
+     * as a failure.
+     *
+     * `fetchSemanticsNodes` **throws** `IllegalStateException: No compose hierarchies found in the
+     * app` when nothing is attached at that instant, and `waitUntil` propagates it on the first
+     * poll instead of waiting out the deadline. This class spends much of its time with another
+     * app in front — the picker, the create-document dialog, the permission dialog — so there is
+     * always a window where the app is coming back and has no composition yet. Before this, that
+     * window was a hard failure: measured on the API 34 leg of run 34057196628, where **both** SAF
+     * tests died that way while the same commit passed API 33, 35, 36 and 37, and the previous
+     * commit passed API 34 and failed 35. A failing leg that moves between runs is #190's
+     * emulator flake, and this is the one place in the class that turned it into a red test.
+     *
+     * **The cost is honest and bounded**: an app that is genuinely gone now fails at the deadline
+     * rather than immediately, so the last composition error is carried into the message to keep
+     * that case diagnosable.
+     */
     private fun awaitNode(tag: String, timeoutMs: Long = APP_TIMEOUT_MS) {
-        composeRule.waitUntil("a node tagged $tag exists", timeoutMs) {
-            composeRule.onAllNodesWithTag(tag).fetchSemanticsNodes().isNotEmpty()
+        var lastError: Throwable? = null
+        try {
+            composeRule.waitUntil("a node tagged $tag exists", timeoutMs) {
+                runCatching { composeRule.onAllNodesWithTag(tag).fetchSemanticsNodes().isNotEmpty() }
+                    .onFailure { lastError = it }
+                    .getOrDefault(false)
+            }
+        } catch (timeout: ComposeTimeoutException) {
+            val note = lastError?.let { "; last composition error: ${it.message}" } ?: ""
+            throw AssertionError("waited ${timeoutMs}ms for a node tagged $tag$note", timeout)
         }
     }
 
@@ -1060,8 +1246,23 @@ class SafPickerRoundTripTest {
         /** Short: either the dialog is up almost immediately, or the permission was already held. */
         const val PERMISSION_DIALOG_MS = 5_000L
 
-        /** A 3 s clip to MP3 on an emulator is about a second; this only bounds a hang. */
-        const val CONVERSION_TIMEOUT_MS = 120_000L
+        /**
+         * Bounds a hang, and **the first number here was measured on one API level and wrong on
+         * another.** It read 120 s, on the strength of the whole test taking 11.8 s on the API 34
+         * CI leg (run 34043502322). API 35 is a different machine: on run 34056545386 the fixture's
+         * `libx265 -crf 24 -preset veryfast` encode started at `20:05:26.897` and the next job in
+         * the suite did not appear until `20:07:41.693` — **134.8 s**, so the encode was still
+         * running when the 120 s bound expired and the test failed with the conversion healthy.
+         *
+         * The logcat is what settles it: `ConversionWorker` logs the route and `FFmpegEngine` the
+         * command, and there is no cancel between them. A timeout that fires on a working
+         * conversion is worse than no bound, because it reads as a product failure.
+         *
+         * 300 s is chosen against that 134.8 s, not against API 34's 11.8 s. **Do not re-tighten
+         * it from a fast leg's timing** — the encode is software on every emulator here, and the
+         * spread between images is larger than any margin a single measurement would suggest.
+         */
+        const val CONVERSION_TIMEOUT_MS = 300_000L
 
         /** The copy is a few kilobytes, but it crosses a provider. */
         const val SAVE_TIMEOUT_MS = 30_000L

docs/e2e-read-findings.md

@@ -382,7 +382,7 @@ decision, not a detail — see **E6** for why no third option exists — and **#
 | **#223** | `HardwareFallbackTest` never attempts the hardware path on any emulator leg | closed — it skips instead of passing vacuously |
 | **#224** | Cancelling a *running* native session, in any of the three engines | closed — all three engines |
 | **#225** | No `content://` input has reached a *successful* conversion — the ffkitsaf bridge | closed, and it found **#238** |
-| **#226** | `OutputPublisher.publish` against a real `DocumentsProvider` | closed — the premise holds; see E7 |
+| **#226** | `OutputPublisher.publish` against a real `DocumentsProvider` | closed — the *premise* holds; see E7. The delete **arm** is still unrun: **#250** |
 | **#227** | The notification's Cancel action has never been fired | closed |
 | **#228** | `encodesFlacLosslessAudio` and `encodesOpus` pass on any non-empty file | closed |
 | **#229** | FFmpeg's progress percentage is computed everywhere and asserted nowhere | closed |
@@ -401,3 +401,117 @@ unasserted value, it was **a combination of two covered things that no test put
 the acceptance criterion wave 4 established and which caught two vacuous tests in that wave before
 they shipped. #223 is the one that shows why the criterion matters: it has two passing assertions and
 still tests nothing.
+
+## The 2026-09-06 re-check
+
+Run after the last ticket landed, to ask whether the suite's self-description had drifted again. It
+had, and **every drifted line came from #226 — the last PR of this read's own wave.**
+
+The suite is 70 tests in 14 classes, 6 carrying `@FailsOnEmulatorApi37`, gating leg 64; the
+committed baseline says 6 and the advisory job agrees (`baseline: matches`). Every gating leg is
+green on `main`.
+
+- **The counts had gone stale in four places** — `CLAUDE.md` (three sites),
+  `FailsOnEmulatorApi37.kt`'s KDoc, and two comments in `status_check.yml` — all still saying five
+  carriers of 69. **The gating figure is what hid it**: 69 − 5 and 70 − 6 are both 64, so the one
+  number a reader would check against a run had not moved. CLAUDE.md's own instruction to derive
+  these rather than remember them is what caught it.
+- **Two KDoc claims in `SafPickerRoundTripTest` described a draft rather than the code.** The save
+  test says MP3 was chosen so the setup could not depend on device codecs; the code converts at the
+  default `MP4_H265` / `FAST`, which routes by `canEncode(H265)`. The *negation* of the stated
+  reason was true. This is **E1 and E3's failure mode landing in a test written by the read that
+  found it** — a passing test with a wrong explanation.
+- **Neither picker test has ever reported on the advisory leg.** The marker's KDoc said the picker
+  test *fails* there behind the rotation test; with six carriers the rotation test truncates the run
+  first, and all four advisory runs at this baseline (`34041156680`, `34041593697`,
+  `34042397320`, `34043502322`) report `expected: 6, received: 4, failed: 4` — the three Media3
+  tests plus the rotation. The save test is therefore
+  marked by **inheritance, not measurement**, which is now what both KDocs say.
+- **One substantive gap, filed as #250.** `FixtureDocumentsProvider.deletedDocumentIds()` has no
+  callers. #226 proved D4's *premise* — SAF hands back a document of exactly zero bytes — but drove
+  only the success path, so `deletePartialOutput` against a real `DocumentsProvider` is still
+  asserted nowhere. `openDestination` is `protected open` precisely to force the failure, so the
+  test is cheap; it costs another marked picker test and a baseline of 7.
+
+**The reusable part is the second bullet.** A read that fixes documentation drift can introduce it in
+the same wave, and the tests it writes are no more self-describing than the ones it audited. The
+check that found it is the one this document already recommends: **read the KDoc against the code,
+not against the ticket.**
+
+## E8 — the instrumented suite's coverage, measured for the first time
+
+**Severity: n/a · Measured 2026-09-06 on API 34 · the number had never existed**
+
+Four coverage waves were steered by a figure that cannot see `app/src/androidTest`. Nothing had
+ever produced the other half, because `enableAndroidTestCoverage` was unset, so a connected run
+emitted no `.ec` at all and `jacocoTestReport`'s execution data names only `testDebugUnitTest`.
+
+Measured by setting that flag temporarily, running `run-e2e.sh 34` (70/70, 0 failed, 3m21s — the
+instrumentation destabilised nothing), and reporting the resulting `.ec` against the **same** class
+directories and exclusions the committed task uses:
+
+| suite | line | branch |
+|---|---|---|
+| JVM `testDebugUnitTest` | 2236/2374 — **94.2%** | 1171/1338 — **87.5%** |
+| Instrumented, 70 tests | 1711/2374 — **72.1%** | 669/1354 — **49.4%** |
+| **Union** | 2342/2374 — **98.7%** | 1212/1354 — **89.5%** |
+
+The JVM row reproduced the committed figure exactly, which is the control: both exec sets match the
+current class files, so the union is trustworthy.
+
+**Two caveats before anyone quotes these.** Branch denominators differ by 16 — 1338 against 1354 —
+entirely inside `MediaProbe`, an artefact of offline versus on-the-fly instrumentation; line
+denominators are identical at 2374, so only the line figures compare exactly. And **72.1% is not a
+grade for the instrumented suite.** Seventy end-to-end tests reach code broadly and choose arms
+rarely; a branch figure of 49.4% is what that shape looks like. This whole document exists because
+the gaps that mattered — #223's vacuous assertions, #238's two covered things nobody combined —
+are invisible to any percentage.
+
+### What the device suite is for, in numbers
+
+It closes **106 lines** the JVM suite misses, and they are precisely the ones wave 4 wrote off:
+
+| file | JVM missed | union missed |
+|---|---|---|
+| `FFmpegEngine.kt` | 32 | **0** |
+| `Media3Engine.kt` | 24 | **0** |
+| `ConcatEngine.kt` | 15 | **0** |
+| `MediaProbe.kt` | 13 | **3** |
+| `MainActivity.kt` | 10 | **1** |
+| `AndroidDeviceCodecs.kt` | 8 | **0** |
+| `Transcoders.kt` | 10 | 3 |
+
+CLAUDE.md's wave-4 read called 81 lines "native or device edges" — `FFmpegEngine` 33,
+`Media3Engine` 24, `ConcatEngine` 14, `MainActivity.onCreate` 10. The first four rows above total
+**81**, and the union leaves **1**. That **confirms** the read's own hypothesis rather than
+overturning it: it always said those zeroes were "the `testDebugUnitTest`-only measurement
+boundary". Nobody had measured past the boundary. `AndroidDeviceCodecs` is the pointed one — #194
+was filed to cut a seam because `probe()` could not be reached, and on a device it is fully covered.
+
+### The 32 lines neither suite reaches, classified
+
+Every one was read. **None of them is an e2e test gap**, which is the result:
+
+| lines | where | classification |
+|---|---|---|
+| 9 | `Transcoders` ×3, `ConversionViewModel`, `ConverterScreen`, `JoinViewModel`, `JoinScreen`, `MainActivity`, `Reattachment` | **compiler-generated** — default-arg `$default` bridges, coroutine completion, the synthetic `NoWhenBranchMatchedException` arm of a `when` over `Destination` |
+| 10 | `ConversionWorker:342-346`, `ConcatWorker:132-136` | `getForegroundInfo()` — WorkManager's **expedited-work** hook, and nothing here enqueues expedited work. The live path is `setForeground(foregroundInfo(...))`, which is covered. **#252** |
+| 3 | `ConversionNotifications:60-62` | **F5** — `areEnabled()` has no callers. Already on record |
+| 3 | `CopyPlanner:28`, `OutputFormat:222-223` | public members with no callers. **#253**, with F5 |
+| 3 | `MediaProbe:210-212` | `probeWithFFprobe`'s `catch` — **F7's sibling, and now measured**. See below |
+| 2 | `FFmpegCommandBuilder:167-168` | `COPY`/`NONE -> error(...)` — F4-shaped, deliberately exempt |
+| 1 | `FFmpegCommandBuilder:188` | the `VORBIS` encode arm. No `OutputFormat` produces it, but `ContainerCapabilities` lists it for WEBM and OGG. **#254** |
+| 1 | `ConversionWorker:231` | `?: error("Could not open the input file.")`. `UnopenableUriTest` fails the job *downstream* of it, so the elvis is unprovoked — F4-shaped, same as the two above |
+
+**`MediaProbe:210-212` is the one that gained a measurement.** F7 ruled `probeWithExtractor`'s catch
+unreachable because Robolectric's `MediaExtractor` never throws. That reasoning does not transfer:
+`probeWithFFprobe` calls `readMediaInformation` in native ffmpeg-kit, which the JVM never loads.
+But `MediaProbe.probe` calls **both** probes on one line, and
+`RemuxTest.probeDistinguishesAudioFromImagesFromRubbish` drives it on a device with 4096 bytes of
+garbage — so the ffprobe path *has* been given malformed input on real hardware and **did not
+throw**. Same conclusion as F7, reached by a different mechanism, and now on record rather than
+assumed.
+
+**The reusable part**: a union report is what separates "no test calls this" from "only a device
+calls it", and neither report alone can. Six of the eight rows above were indistinguishable from
+real gaps in the JVM-only number.

tools/git-hooks/local-gate.sh

@@ -0,0 +1,165 @@
+#!/usr/bin/env bash
+#
+# The local gate: what has to be green before a commit is made or a branch is pushed.
+#
+# THE RULE THIS ENFORCES (2026-09-06). Source changes must have the unit tests AND the
+# instrumented tests passing at every supported API level before they are committed or
+# pushed; test changes must have the whole suite passing at every API level. CI is not the
+# place to find out. Four legs of this repo's history were spent discovering on CI what a
+# local sweep would have said in twenty minutes -- and worse, the failing leg MOVED between
+# runs (API 35 red then green, API 34 green then red), which is exactly the signal that gets
+# misread as "someone else's flake" when it is read one leg at a time.
+#
+# WHY BOTH HOOKS RUN THE SAME GATE. A pre-commit-only gate is bypassed by amending; a
+# pre-push-only gate lets a broken commit exist locally and get rebased into something else.
+# Running both is not redundant in practice because of the cache below.
+#
+# THE CACHE IS KEYED ON CONTENT, NOT ON TIME, AND ON THE RIGHT CONTENT. The sweep is recorded
+# under the hash of the `app/src` SUBTREE it verified, not the whole repo tree. Keying it on the
+# whole tree was the first cut and it was wrong in a way that would have trained people to hate
+# this hook: editing a comment in CLAUDE.md, or in this script, invalidated a sweep of identical
+# application code and re-ran forty minutes of emulators to prove nothing. What the sweep is
+# evidence about is `app/src`; that is what it is filed under. Any change to a single byte under
+# `app/src` still invalidates it. The JVM gate is cheap and runs unconditionally.
+#
+# WHAT COUNTS AS "EVERY SUPPORTED API LEVEL", AND WHY 37 IS NOT AN EMULATOR HERE. 33, 34, 35
+# and 36 run the whole suite on emulators. **API 37 cannot be run on an emulator on this host at
+# all** -- not "is red", cannot run: measured 2026-09-06, the image logs
+# `3 new surfaceflinger aborts in 45 s (want 0)` and then the APK install itself fails with
+# `Can't find service: package`, because the framework is already gone before Gradle gets to
+# install anything. `Starting 0 tests`. That is the same gralloc abort docs/api-37-emulator-crash.md
+# measures, hit earlier in the sequence than the suite.
+#
+# So API 37 is covered here by the physical Pixel 10 Pro XL when it is attached, and by CI's
+# gating leg otherwise. The hook says loudly which of the two happened rather than quietly
+# claiming five levels when it ran four.
+#
+# THERE IS DELIBERATELY NO SKIP VARIABLE. An `LMC_SKIP_E2E=1` would be `--no-verify` wearing
+# a different hat, and `--no-verify` needs the repo owner's say-so each time. If this gate is
+# wrong, fix the gate.
+set -uo pipefail
+
+REPO_ROOT="$(git rev-parse --show-toplevel)"
+cd "$REPO_ROOT" || exit 1
+
+MODE="$(basename "$0")"
+ZERO="0000000000000000000000000000000000000000"
+CACHE_DIR=".git/lmc-verify"
+GRADLE_GATE=(:app:assembleDebug :app:testDebugUnitTest :app:compileDebugAndroidTestKotlin
+  :app:ktlintCheck :app:detekt :app:lintDebug)
+
+say() { printf '\n\033[1m[local-gate]\033[0m %s\n' "$*"; }
+die() {
+  printf '\n\033[1;31m[local-gate] BLOCKED\033[0m %s\n' "$*"
+  printf '  The rule: source work needs unit + e2e green at every API level before commit/push;\n'
+  printf '  test work needs the whole suite green at every level. Fix it, or ask before using\n'
+  printf '  --no-verify -- that flag is not yours to reach for unprompted.\n\n'
+  exit 1
+}
+
+# --- what changed, and what tree is being verified ---------------------------------------
+
+changed_files=""
+tree=""
+case "$MODE" in
+  pre-commit)
+    changed_files="$(git diff --cached --name-only --diff-filter=ACMR)"
+    tree="$(git rev-parse "$(git write-tree):app/src" 2>/dev/null || echo "")"
+    ;;
+  pre-push)
+    # stdin is `<local ref> <local sha> <remote ref> <remote sha>`, one line per ref pushed.
+    while read -r _ local_sha _ remote_sha; do
+      [ "$local_sha" = "$ZERO" ] && continue          # branch deletion carries no content
+      base="$remote_sha"
+      if [ "$remote_sha" = "$ZERO" ]; then
+        # A new branch: compare against main rather than against every commit ever made.
+        base="$(git merge-base origin/main "$local_sha" 2>/dev/null || echo "")"
+      fi
+      if [ -n "$base" ]; then
+        changed_files="$changed_files$(git diff --name-only --diff-filter=ACMR "$base" "$local_sha")"$'\n'
+      else
+        changed_files="$changed_files$(git show --pretty=format: --name-only "$local_sha")"$'\n'
+      fi
+      tree="$(git rev-parse "$local_sha:app/src" 2>/dev/null || echo "")"
+    done
+    ;;
+  *)
+    say "unknown hook name '$MODE'; nothing to do"
+    exit 0
+    ;;
+esac
+
+if [ -z "${changed_files//[[:space:]]/}" ]; then
+  say "no added/modified files; nothing to verify"
+  exit 0
+fi
+
+touches_source=0
+touches_tests=0
+while IFS= read -r f; do
+  case "$f" in
+    app/src/main/*) touches_source=1 ;;
+    app/src/test/*|app/src/androidTest/*) touches_tests=1 ;;
+  esac
+done <<< "$changed_files"
+
+# --- the cheap gate always runs -----------------------------------------------------------
+
+say "$MODE: running the JVM gate"
+if ! ./gradlew "${GRADLE_GATE[@]}" --continue; then
+  die "the JVM gate failed (assemble, unit tests, androidTest compile, ktlint, detekt, lint)."
+fi
+
+# --- the sweep, when code is involved ------------------------------------------------------
+
+if [ "$touches_source" -eq 0 ] && [ "$touches_tests" -eq 0 ]; then
+  say "no app/src changes; the instrumented sweep is not required for this one"
+  mkdir -p "$CACHE_DIR" && [ -n "$tree" ] && : > "$CACHE_DIR/$tree"
+  exit 0
+fi
+
+if [ -n "$tree" ] && [ -f "$CACHE_DIR/$tree" ]; then
+  say "app/src ($tree) already swept and green; nothing under app/src has changed since"
+  exit 0
+fi
+
+say "app/src changed -- sweeping API 33, 34, 35, 36 (this takes tens of minutes, by design)"
+if ! tools/local-emulator/run-e2e.sh 33 34 35 36; then
+  die "the instrumented suite is not green on 33-36."
+fi
+
+# API 37: the physical device if it is here, and an honest statement if it is not. run-e2e.sh is
+# emulator-only (and overwrites E2E_EXTRA_GRADLE_ARGS with --rerun, so extra args cannot be passed
+# through it), so this drives Gradle directly with the serial pinned -- the phone must never be
+# picked up by accident, which is the hazard run-e2e.sh's header calls out.
+export ANDROID_HOME="${ANDROID_HOME:-$HOME/Android/Sdk}"
+export PATH="$ANDROID_HOME/platform-tools:$PATH"
+
+device=""
+while read -r serial state; do
+  [ "$state" = "device" ] || continue
+  case "$serial" in emulator-*) continue ;; esac
+  [ "$(adb -s "$serial" shell getprop ro.build.version.sdk 2>/dev/null | tr -d '\r')" = "37" ] || continue
+  device="$serial"
+  break
+done < <(adb devices 2>/dev/null | tail -n +2)
+
+levels="33, 34, 35, 36"
+if [ -n "$device" ]; then
+  say "API 37 on the attached device $device"
+  if ! ANDROID_SERIAL="$device" ./gradlew :app:connectedDebugAndroidTest -PabiFilters=arm64-v8a; then
+    die "the instrumented suite is not green on API 37 (device $device)."
+  fi
+  levels="$levels, 37"
+else
+  say "NOT COVERED LOCALLY: API 37. No API 37 device is attached, and the API 37 emulator cannot
+  install the APK on this host (see this script's header). CI's gating leg is what answers for it;
+  attach the Pixel 10 Pro XL to have this hook cover it too."
+fi
+
+mkdir -p "$CACHE_DIR" && [ -n "$tree" ] && : > "$CACHE_DIR/$tree"
+# Name the levels rather than claiming "every supported level". The first cut said the latter on
+# both paths, including the one that had just printed NOT COVERED two lines above -- a false claim
+# printed by the tool whose whole job is to stop false claims reaching CI.
+say "green on API $levels; $MODE allowed"
+exit 0

tools/git-hooks/pre-commit

@@ -0,0 +1 @@
+local-gate.sh

tools/git-hooks/pre-push

@@ -0,0 +1 @@
+local-gate.sh