ADFA-4128 (11/11): app + bench — wiring Quick Build into the IDE and the benchmark harness - #1723

…ent, test gaps

Important 1 (daemon idle-timeout/Metaspace tuner un-gated): kept un-gated by
design — the 384m Metaspace floor fixes real OOM-killed builds and the tiered
idle timeouts keep low-RAM devices from losing the IDE to lmkd; GradleBuildTuner
now states this in its KDoc. Ships flag-off; needs Bryan sign-off in the PR body.

Important 2 (generateSources narrowing un-gated): judged a genuine all-users
improvement, not QB-specific — the old code ran a Gradle generateSources after
EVERY save-all (and after any XML save in SaveFileAction), a per-save build tax;
flag-off the deferral degenerates to the same immediate call, so the narrowing
is the only behavior change. Known trade (manifest-only edits leave generated
Manifest/R intermediates stale until the next resource save or build) now
stated at both call sites. Ships flag-off; needs sign-off in the PR body.

Important 3 (install dropped on rotation, flag on): installApk's async path now
re-arms AwaitingInstall (BuildViewModel.reArmInstall, fires only from Idle)
from the coroutine's drop path, so a configuration change during the APK-manifest
parse makes the recreated activity's collector retry the install instead of
silently losing a successful build. Covered by BuildViewModelInstallReArmTest.

Test gap (zip-slip guard): extraction loop extracted to
QuickBuildArtifactStager.extractDaemonZip(InputStream, File); the guard is
watched going red by QuickBuildArtifactStagerTest (a ../ entry throws and
nothing lands outside the daemon dir).

Test gap (InstallationEventFlow mapping): InstallationEventFlowTest pins the
PackageInstaller status mapping, including the ABORTED-vs-FAILURE branch order
and the no-extras / no-status paths.

Test gap (service-side output capture): suppress/capture/drain routing
extracted from GradleBuildService.logOutput into InternalBuildOutputCapture;
bounded tail, drain-clears, throwing progress listener and editor-listener
routing pinned by InternalBuildOutputCaptureTest.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Kj9YeCDHGp9DU8LPtfWJ7W

…tiple

The Gradle daemon idle-timeout comment quoted a speedup multiple, which put a
benchmark figure into shipping production code. The reason the timeout is
generous is structural - a warm daemon skips the cold start, which dominates a
short rebuild - so the comment now says that instead.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Kj9YeCDHGp9DU8LPtfWJ7W

- F1723-1 put QuickBuildPipelineTest into the suite that actually runs
- F1723-4 guard the deferred prebuild fire() against a throw

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FstXxJ5cwWPcvmhZ9vJgJ7

Six places where a Quick Build doc contradicted the code it describes. Each was
re-verified against the source rather than taken from the review comment.

- debugging.md: the deploy round-trip row said the 15 s bound covers "one AIDL
  onPayload call". IQuickBuildTarget is a oneway interface, so that call returns
  immediately; DeployChannel wraps the call plus the wait for a
  generation-matched report, which is what its own KDoc already said.
- why-not-android-jar.md: listed native libs as hot-loadable. A .so under
  jniLibs forces a Gradle fallback (ChangeClassifier). Loadable at runtime and
  changeable via live reload are different properties.
- reliability-gaps.md: "five user-facing defects" against three fixed and four
  open. Seven were surfaced; the fixed three are relink-stuck, #88 and #90. Also
  states why Blocks v1? reads TBD - the decision at the top is a proposal, and
  the cells become "No" when it is confirmed.
- low-spec-devices.md: stated an inferred mechanism (SerialGC thrashing in a
  small heap) as the confirmed cause of the 1.9 GB failure. The outcome is
  measured; the mechanism is not, and the uncapped run that would confirm it is
  still unmeasured. Retitled to what was actually observed.
- concurrency.md: the tap-races-its-own-save section read as current behaviour.
  It describes the pre-2026-08-13 design that the redesign below it replaced.
- perf-roadmap.md: incomplete sentence.

Not applied: CodeRabbit's finding that manual-qa.md's screenrecord
--time-limit 1740 is invalid because AOSP caps at 180 s. False on our hardware -
recordings of 1774 s, 2432 s, 2592 s, 2842 s and 3534 s have all completed on
the A56, and the surrounding comment already documents the real 30-minute cap
that 1740 sits under. Applying it would break working recordings.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FstXxJ5cwWPcvmhZ9vJgJ7

None of these is a Quick Build failure - the walk passed every test. They are
places where the product is correct and unhelpful.

A2 - the bolt read identically to a screen reader in READY and ERROR. Every tone
has its own icon shape, so a sighted user can tell them apart; collapsing them
all to "Quick Build" hid that distinction from exactly the user who cannot see
the icon. ERROR, SLOW and RECONNECTING now announce their state. BUILDING and
the standard-build-blocked case already did.

A3 - after an undeliverable build the bar read "built, but could not be
delivered - see Build Output" on every poll, while the sentence naming the fix
("Your app is not running. Tap Quick Build to start it with your changes.") was
only in Build Output. The bar now names the tap when that is the whole problem.
Carried as a typed flag rather than matched on the message text, the same way
proxyAppNotConnected already is, and kept separate from it because they mean
opposite things: appNotRunning is "nobody opened it", proxyAppNotConnected is
"we launched it and it still did not arrive".

A4 - an orphaned proxy app reported CoGo's expected connect() rejection at W on
every attempt of the rebind backoff loop, 14 times in one restart window. The
behaviour is right (it continues standalone); repeating an expected rejection at
W buries the entries around it. Reported once per streak now, cleared by a
successful connect.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FstXxJ5cwWPcvmhZ9vJgJ7

The page proposed that #87, #89, #91 and the relink-crash gap go to v1.1,
then left the table's "Blocks v1?" column reading TBD on all four rows. A
proposal in a title and a TBD in a table say different things to a reader,
and CodeRabbit flagged the pair as an internal inconsistency.

Decision confirmed 2026-08-25: none of the four block v1. The four cells now
read "No - v1.1", the title states the answer rather than asking it, and the
prose no longer describes itself as awaiting confirmation.

No change to any gap's evidence, root cause, or fix.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FstXxJ5cwWPcvmhZ9vJgJ7

…ing branch

F1713-3 (docs/concurrency.md): the thesis said every expensive thing runs
in another process while the table directly below it put the mtime poll
and the install call on Dispatchers.IO inside CoGo. Name the exception.

F1713-8 (docs/manual-qa.md): files are not killed, processes are - and a
teammate follows this runbook literally while holding a half-recorded QA
session. Say screenrecord.

Both patch text that exists only in the four trailing commits, so they
could not ship until those commits had a home.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xsc7AMGBVyEMfrwpZX87iC