…te (QVAC-18168 follow-up)

Rebased clean on main after PR #1913 merged. Each monolithic mobile
workflow (~1400-1800 lines) replaced with a thin composite-based shim
(~170-230 lines).

Addons migrated:
  embed-llamacpp, bci-whispercpp, transcription-whispercpp,
  transcription-parakeet, decoder-audio, diffusion-cpp,
  classification-ggml, tts-onnx (q4/q4f16 variant matrix), tts-ggml

Composite extensions (backwards-compatible, no change for LLM/OCR/NMT):
  - setup: skip-prebuilds input (decoder-audio has no own prebuilds)
  - monitor: max-wait-time-seconds input (tts-onnx needs 3h)

Addon-side provision scripts (matching NMT's pattern):
  - packages/tts-ggml/scripts/provision-mobile-models.sh
  - packages/transcription-parakeet/scripts/provision-mobile-models.sh

Runner alignment: all shims use qvac-ubuntu2404-x64 for Android
(matching main's latest self-hosted strategy from PR #2021/#2123).

Co-authored-by: Cursor <cursoragent@cursor.com>

…omposite (#2153)

* refactor(mobile-test): migrate remaining 9 addons onto shared composite (QVAC-18168 follow-up)

Rebased clean on main after PR #1913 merged. Each monolithic mobile
workflow (~1400-1800 lines) replaced with a thin composite-based shim
(~170-230 lines).

Addons migrated:
  embed-llamacpp, bci-whispercpp, transcription-whispercpp,
  transcription-parakeet, decoder-audio, diffusion-cpp,
  classification-ggml, tts-onnx (q4/q4f16 variant matrix), tts-ggml

Composite extensions (backwards-compatible, no change for LLM/OCR/NMT):
  - setup: skip-prebuilds input (decoder-audio has no own prebuilds)
  - monitor: max-wait-time-seconds input (tts-onnx needs 3h)

Addon-side provision scripts (matching NMT's pattern):
  - packages/tts-ggml/scripts/provision-mobile-models.sh
  - packages/transcription-parakeet/scripts/provision-mobile-models.sh

Runner alignment: all shims use qvac-ubuntu2404-x64 for Android
(matching main's latest self-hosted strategy from PR #2021/#2123).

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-test): parakeet — match monolith's mocha/WDIO timeouts (45min / 10min)

Main monolith uses timeout: 2700000 (45min) and
waitforTimeout: 600000 (10min). Our composite defaults to 1800000
(30min) and 120000 (2min). The slower parakeet tests (sortformer
inference on Pixel 9a) exceed 30min and time out.

Pass mocha-timeout-ms: 2700000 and wdio-waitfor-timeout-ms: 600000
to upload-to-devicefarm to match the monolith.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-test): match monolith mocha/WDIO timeouts for 4 addons + add tts-onnx perf filter

Deep audit of all 9 monoliths revealed custom timeout values that
our shims were missing (using composite defaults instead):

  bci-whispercpp:          mocha 900000 (15min), was 1800000
  transcription-whispercpp: mocha 900000 (15min), was 1800000
  decoder-audio:           mocha 600000 (10min), was 1800000
  tts-ggml:                mocha 2700000 (45min) + waitfor 600000 (10min)

Also: tts-onnx monolith used --filter supertonic on perf extraction
to exclude Chatterbox rows from reports. Added filter: 'supertonic'
to the extract-addon-perf call.

embed-llamacpp, diffusion-cpp, classification-ggml, tts-onnx all
matched the composite defaults (1800000 / 120000) — no change needed.
transcription-parakeet was already fixed in the previous commit.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-setup): always try artifact download, fall back to npm only when empty

The artifact-download steps were gated behind
github.event_name != 'workflow_dispatch', which skipped them on
workflow_dispatch even when on-pr-* had just produced fresh
prebuild artifacts in sibling jobs. This caused workflow_dispatch
runs to always fall back to npm, getting outdated/smaller prebuilds
(e.g. parakeet 20 MB from npm vs 68 MB from fresh artifacts).

Fix: remove the event_name gate from artifact download (with
continue-on-error: true it's safe to run when no artifacts exist).
The npm-fallback step now checks if prebuilds/ already has content
from artifacts before attempting npm pack.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-schedule): bump default Device Farm jobTimeoutMinutes from 60 to 90

Pixel 9 Pro runs LLM VLM inference ~1.7x slower than Samsung S25/S26
Ultra. The groupImagesPerf shard takes ~56 min on Pixel, and Device
Farm's 60-min job timeout STOPS the run during teardown even though
all 3 tests passed. Bumping to 90 min gives enough headroom.

NMT already overrides to 120 via the consumer shim.

Co-authored-by: Cursor <cursoragent@cursor.com>

* feat(mobile-monitor): distinguish STOPPED-but-passed from real failures

Device Farm result=STOPPED means the jobTimeoutMinutes cap expired,
not that tests failed. When a device is STOPPED but its counters
show 0 failed / 0 errored / N passed, the tests all completed
successfully — DF just killed the teardown phase.

Before: STOPPED counted as USER_FAILED, triggering exit 1 even
though every test passed. This burned investigation time.

Now: STOPPED with clean counters → ⚠️ warning + USER_PASSED.
STOPPED with actual failures → ❌ with counter breakdown.
WARNED → treated as success (same as PASSED).
FAILED / ERRORED → ❌ with counter breakdown.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-upload): implement perf-only group filtering via perf-test-regex input

The monolith had an inline filter_perf() + maybe_make_and_upload()
that intersected each test group's grep with a perf-test regex when
qvac_perf_only=true, skipping groups with no matching perf tests.
This was lost when the composite was created — qvac-perf-only was
threaded through to on-device config but the scheduling-side filter
was missing. Result: benchmark runs scheduled ALL test groups on
ALL devices instead of the perf-emitting subset.

New perf-test-regex input on upload-to-devicefarm: when
qvac-perf-only=true and perf-test-regex is set, each group's grep
is filtered to only keep matching tests. Empty groups are skipped
with a clear log message.

LLM consumer now passes the same PERF_REGEX the monolith used:
  ^(runImageElephantTest|runImageFruitPlateTest|runImageHighResAuroraTest|runBitnetTest|runToolCallingTest)$

Other addons don't use qvac_perf_only so they're unaffected.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: move perf-test filter regex to consumer input instead of test-groups.json

The perf_tests key in test-groups.json broke the LLM addon's
generate-mobile-integration-tests.js validator, which treats every
top-level key as a platform and expects all tests to be covered.

Match the original monolith approach: the perf-emitting test regex
is supplied by the consumer workflow via a new `perf-test-regex`
composite input, keeping test-groups.json identical to main.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: convention-based perf-test filtering via perf-tests.json

Replace the hardcoded perf-test-regex consumer input with a
convention-based auto-discovery file (perf-tests.json) that sits
alongside test-groups.json in the addon's test/mobile/ directory.

The composite reads the file when qvac_perf_only=true, builds the
filter regex from the array, and skips groups with no perf-emitting
tests. No consumer workflow changes needed — addons opt in by
dropping a perf-tests.json file.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Cursor <cursoragent@cursor.com>

…#2123)

…omposite (#2153)

* refactor(mobile-test): migrate remaining 9 addons onto shared composite (QVAC-18168 follow-up)

Rebased clean on main after PR #1913 merged. Each monolithic mobile
workflow (~1400-1800 lines) replaced with a thin composite-based shim
(~170-230 lines).

Addons migrated:
  embed-llamacpp, bci-whispercpp, transcription-whispercpp,
  transcription-parakeet, decoder-audio, diffusion-cpp,
  classification-ggml, tts-onnx (q4/q4f16 variant matrix), tts-ggml

Composite extensions (backwards-compatible, no change for LLM/OCR/NMT):
  - setup: skip-prebuilds input (decoder-audio has no own prebuilds)
  - monitor: max-wait-time-seconds input (tts-onnx needs 3h)

Addon-side provision scripts (matching NMT's pattern):
  - packages/tts-ggml/scripts/provision-mobile-models.sh
  - packages/transcription-parakeet/scripts/provision-mobile-models.sh

Runner alignment: all shims use qvac-ubuntu2404-x64 for Android
(matching main's latest self-hosted strategy from PR #2021/#2123).

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-test): parakeet — match monolith's mocha/WDIO timeouts (45min / 10min)

Main monolith uses timeout: 2700000 (45min) and
waitforTimeout: 600000 (10min). Our composite defaults to 1800000
(30min) and 120000 (2min). The slower parakeet tests (sortformer
inference on Pixel 9a) exceed 30min and time out.

Pass mocha-timeout-ms: 2700000 and wdio-waitfor-timeout-ms: 600000
to upload-to-devicefarm to match the monolith.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-test): match monolith mocha/WDIO timeouts for 4 addons + add tts-onnx perf filter

Deep audit of all 9 monoliths revealed custom timeout values that
our shims were missing (using composite defaults instead):

  bci-whispercpp:          mocha 900000 (15min), was 1800000
  transcription-whispercpp: mocha 900000 (15min), was 1800000
  decoder-audio:           mocha 600000 (10min), was 1800000
  tts-ggml:                mocha 2700000 (45min) + waitfor 600000 (10min)

Also: tts-onnx monolith used --filter supertonic on perf extraction
to exclude Chatterbox rows from reports. Added filter: 'supertonic'
to the extract-addon-perf call.

embed-llamacpp, diffusion-cpp, classification-ggml, tts-onnx all
matched the composite defaults (1800000 / 120000) — no change needed.
transcription-parakeet was already fixed in the previous commit.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-setup): always try artifact download, fall back to npm only when empty

The artifact-download steps were gated behind
github.event_name != 'workflow_dispatch', which skipped them on
workflow_dispatch even when on-pr-* had just produced fresh
prebuild artifacts in sibling jobs. This caused workflow_dispatch
runs to always fall back to npm, getting outdated/smaller prebuilds
(e.g. parakeet 20 MB from npm vs 68 MB from fresh artifacts).

Fix: remove the event_name gate from artifact download (with
continue-on-error: true it's safe to run when no artifacts exist).
The npm-fallback step now checks if prebuilds/ already has content
from artifacts before attempting npm pack.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-schedule): bump default Device Farm jobTimeoutMinutes from 60 to 90

Pixel 9 Pro runs LLM VLM inference ~1.7x slower than Samsung S25/S26
Ultra. The groupImagesPerf shard takes ~56 min on Pixel, and Device
Farm's 60-min job timeout STOPS the run during teardown even though
all 3 tests passed. Bumping to 90 min gives enough headroom.

NMT already overrides to 120 via the consumer shim.

Co-authored-by: Cursor <cursoragent@cursor.com>

* feat(mobile-monitor): distinguish STOPPED-but-passed from real failures

Device Farm result=STOPPED means the jobTimeoutMinutes cap expired,
not that tests failed. When a device is STOPPED but its counters
show 0 failed / 0 errored / N passed, the tests all completed
successfully — DF just killed the teardown phase.

Before: STOPPED counted as USER_FAILED, triggering exit 1 even
though every test passed. This burned investigation time.

Now: STOPPED with clean counters → ⚠️ warning + USER_PASSED.
STOPPED with actual failures → ❌ with counter breakdown.
WARNED → treated as success (same as PASSED).
FAILED / ERRORED → ❌ with counter breakdown.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(mobile-upload): implement perf-only group filtering via perf-test-regex input

The monolith had an inline filter_perf() + maybe_make_and_upload()
that intersected each test group's grep with a perf-test regex when
qvac_perf_only=true, skipping groups with no matching perf tests.
This was lost when the composite was created — qvac-perf-only was
threaded through to on-device config but the scheduling-side filter
was missing. Result: benchmark runs scheduled ALL test groups on
ALL devices instead of the perf-emitting subset.

New perf-test-regex input on upload-to-devicefarm: when
qvac-perf-only=true and perf-test-regex is set, each group's grep
is filtered to only keep matching tests. Empty groups are skipped
with a clear log message.

LLM consumer now passes the same PERF_REGEX the monolith used:
  ^(runImageElephantTest|runImageFruitPlateTest|runImageHighResAuroraTest|runBitnetTest|runToolCallingTest)$

Other addons don't use qvac_perf_only so they're unaffected.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: move perf-test filter regex to consumer input instead of test-groups.json

The perf_tests key in test-groups.json broke the LLM addon's
generate-mobile-integration-tests.js validator, which treats every
top-level key as a platform and expects all tests to be covered.

Match the original monolith approach: the perf-emitting test regex
is supplied by the consumer workflow via a new `perf-test-regex`
composite input, keeping test-groups.json identical to main.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: convention-based perf-test filtering via perf-tests.json

Replace the hardcoded perf-test-regex consumer input with a
convention-based auto-discovery file (perf-tests.json) that sits
alongside test-groups.json in the addon's test/mobile/ directory.

The composite reads the file when qvac_perf_only=true, builds the
filter regex from the array, and skips groups with no perf-emitting
tests. No consumer workflow changes needed — addons opt in by
dropping a perf-tests.json file.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Cursor <cursoragent@cursor.com>