The animation/transition test suite (added in #4821) creates 7 mutable
images per test × ~17 tests, exhausting Metal device memory mid-suite
and hanging the runner after the third test. Root cause: GLUIImage's
Metal ivars and CN1Metalcompat's lazy texture caches all held +1 retains
that no one was releasing.

Fixes:
- GLUIImage.dealloc releases mtlTexture, mtlMutableTexture and
  mtlMutableCommandBuffer (was nil-out only).
- setMtlMutableTexture / setMtlMutableCommandBuffer now retain incoming
  values and release the previous slot. Without retain the autoreleased
  command buffer from [queue commandBuffer] would dangle after the next
  pool drain.
- setImage: releases the cached mtlTexture before nil-ing it.
- CN1MetalEnsureMutableTexture releases the +1 from
  newTextureWithDescriptor: now that the setter retains.
- CN1MetalReadMutableImagePixels releases the staging "shared" texture
  after copy-out (previously leaked one full-resolution texture per
  Image.getRGB call).
- Text and gradient caches release the evicted slot's +1 retains before
  overwriting (text cache also evicts a leaked NSString key per slot
  rotation).

Same MRR ownership pattern as CN1MetalGlyphAtlas (b9c5add / 4de8cb0).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Reinstates the drain-and-wait + Metal blit path disabled in 9b2aaf1
("disable step-6 readback (suspected hang in DrawImage)"). The legacy
fallback returns the stale UIImage initial-fill for any mutable that's
been drawn into post-Phase 3, which was tolerable until master #4821
landed the animation/transition test suite — those tests build screenshots
by reading back from per-frame mutables via Image.getRGB and PNG-encode,
so the disabled path means every animation grid emits empty pixels and
downstream EDT scheduling stalls on the empty-result follow-up.

Original GPU-pressure deadlock concern (nextDrawable wedge) is mitigated
by the resource-leak fix in e548d1a: GLUIImage no longer pins +1
retains on its mtlMutableTexture / mutable command buffer / shared
readback texture, so transient mutables can release back to the device
between frames.

Bracketed with finishDrawingOnImageImpl so an in-progress draw session
is closed before the drain (queued ops drained via flushBuffer's
dispatch_sync(main, drawFrame)) and reopened after the blit-and-read,
mirroring the legacy fallback's pattern.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

PNG/JPEG encoding flows through ImageIO.save → createImageFile, which
calls [GLUIImage getImage] and encodes the original UIImage backing.
For any mutable image that has been drawn into via the Phase 3 v2 Metal
pipeline that backing is just the initial fill — so emitImage on the
animation/transition tests added in master #4821 was producing black
grids regardless of what got painted into the frames.

Drain the op queue (outside the dispatch_sync wrapper, to avoid nested
dispatch_sync to main) so the mutable's MTLTexture is up-to-date, then
inside the dispatch_sync block blit-and-read those pixels into a fresh
UIImage and encode that. Matches the imageRgbToIntArrayImpl path
re-enabled in the previous commit.

If the JNI is invoked while a draw-on-image session is open against
this peer, finishDrawing is called before the drain and reopened
afterwards so the surrounding Java code observes a continuous session.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ridge

PR #4821 added 124 lines to Cn1ssDeviceRunner.java, which shifted the
translator's lambda numbering: `lambda_awaitTestCompletion_3_*` became
`_0_`, and `lambda_finalizeTest_4_*` became `_0_`. port.js still
referenced the old `_3_` / `_4_` IDs hardcoded.

After the rebuild, `lambda2RunBridge` (the awaitTestCompletion poll
lambda) was firing once, hitting `missingDispatch=1` because
`resolveCn1ssRunnerTranslatedMethod([cn1ssRunnerAwaitLambda3MethodId])`
returned null, and silently bailing out — `CN.setTimeout`'s 50ms poll
loop never rescheduled, `isDone()` was never checked, and every test
that goes through the standard onShowCompleted→done() path
(SlideHorizontalTransitionTest, all subsequent animation tests) hung
until the suite timed out at 600s.

Smoking gun in the browser log:

  lambda3RunBridge:dispatch:index=1:nextIndex=2     (MainScreen finalized)
  lambda2RunBridge:HIT                              (Slide poll fires once)
  lambda2RunBridge:missingDispatch=1                (lookup fails, polling dies)

Build the candidate-id list by looping 0..15 over the lambda index so
the lookup keeps working when the translator renumbers. Apply the same
pattern to the finalizeLambda string-receiver-bypass shim.

Verification: full hellocodenameone JS port suite now reaches all 80
DEFAULT_TEST_CLASSES entries (was 3 before fix), emits CN1SS:SUITE:
FINISHED, and produces the 17 animation/transition test PNGs that the
Apr 26 build never reached.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ridge

PR #4821 added 124 lines to Cn1ssDeviceRunner.java, which shifted the
translator's lambda numbering: `lambda_awaitTestCompletion_3_*` became
`_0_`, and `lambda_finalizeTest_4_*` became `_0_`. port.js still
referenced the old `_3_` / `_4_` IDs hardcoded.

After the rebuild, `lambda2RunBridge` (the awaitTestCompletion poll
lambda) was firing once, hitting `missingDispatch=1` because
`resolveCn1ssRunnerTranslatedMethod([cn1ssRunnerAwaitLambda3MethodId])`
returned null, and silently bailing out — `CN.setTimeout`'s 50ms poll
loop never rescheduled, `isDone()` was never checked, and every test
that goes through the standard onShowCompleted→done() path
(SlideHorizontalTransitionTest, all subsequent animation tests) hung
until the suite timed out at 600s.

Smoking gun in the browser log:

  lambda3RunBridge:dispatch:index=1:nextIndex=2     (MainScreen finalized)
  lambda2RunBridge:HIT                              (Slide poll fires once)
  lambda2RunBridge:missingDispatch=1                (lookup fails, polling dies)

Build the candidate-id list by looping 0..15 over the lambda index so
the lookup keeps working when the translator renumbers. Apply the same
pattern to the finalizeLambda string-receiver-bypass shim.

Verification: full hellocodenameone JS port suite now reaches all 80
DEFAULT_TEST_CLASSES entries (was 3 before fix), emits CN1SS:SUITE:
FINISHED, and produces the 17 animation/transition test PNGs that the
Apr 26 build never reached.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ridge

PR #4821 added 124 lines to Cn1ssDeviceRunner.java, which shifted the
translator's lambda numbering: `lambda_awaitTestCompletion_3_*` became
`_0_`, and `lambda_finalizeTest_4_*` became `_0_`. port.js still
referenced the old `_3_` / `_4_` IDs hardcoded.

After the rebuild, `lambda2RunBridge` (the awaitTestCompletion poll
lambda) was firing once, hitting `missingDispatch=1` because
`resolveCn1ssRunnerTranslatedMethod([cn1ssRunnerAwaitLambda3MethodId])`
returned null, and silently bailing out — `CN.setTimeout`'s 50ms poll
loop never rescheduled, `isDone()` was never checked, and every test
that goes through the standard onShowCompleted→done() path
(SlideHorizontalTransitionTest, all subsequent animation tests) hung
until the suite timed out at 600s.

Smoking gun in the browser log:

  lambda3RunBridge:dispatch:index=1:nextIndex=2     (MainScreen finalized)
  lambda2RunBridge:HIT                              (Slide poll fires once)
  lambda2RunBridge:missingDispatch=1                (lookup fails, polling dies)

Build the candidate-id list by looping 0..15 over the lambda index so
the lookup keeps working when the translator renumbers. Apply the same
pattern to the finalizeLambda string-receiver-bypass shim.

Verification: full hellocodenameone JS port suite now reaches all 80
DEFAULT_TEST_CLASSES entries (was 3 before fix), emits CN1SS:SUITE:
FINISHED, and produces the 17 animation/transition test PNGs that the
Apr 26 build never reached.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>