[leehack]

Summary

• Closes four test gaps from the #1837 audit follow-ups (Require the in-process LiteRT marker for Linux and Windows LiteRT-LM embeddings #1830, Refuse LiteRT-LM GPU on Linux when every Vulkan device is software #1831, 0.19.1 batch: #1830 #1831 #1833 #1839 #1840 #1841 #1842 #1843); test-only, no production change.

Production-readiness scope

• User-facing scope: None; tests only.
• Supported platforms/paths: VM tests on macOS and Linux.
• Unsupported or intentionally unavailable paths: N/A.
• Out of scope / follow-ups: None from this PR. The remaining 0.19.1 follow-ups from audits #1837 items (bundled-runtime marker export probe, coverage-lib row) are other lanes of #1845.

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly. (N/A, test-only)
• Docs/changelog. (N/A, test-only)
• Regression coverage covers the original issue plus key negative paths.
• Validation coverage registry. (N/A)
• Security/privacy review completed: no secrets or paths leak.
• Follow-up work is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Bugfix PR: N/A, test-only; no production change.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• Touched suites (litert_lm_v1_embedding, litert_lm_gpu_adapter, litert_lm_service, litert_lm_embedding_native) plus test/unit/tooling/package_skills_test.dart and test/unit/test_structure: all pass (-j 1, isolated TMPDIR)
• Full dart test -p vm -j 1 and chrome: N/A, CI runs them
• coverage-lib: N/A, no lib/ change

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
N/A	Test-only LiteRT-LM fixtures	macOS arm64 VM, compiled C fixtures	Pass	mutations below

Mutation results (each reverted):

#	Mutation	Result
1	LiteRtLmRuntimeClient._selectRuntime passes marker 1 when _facadeLibraryPath == null (env-dir route)	new env-dir refusal tests fail (3 failures: linux_arm64, linux_x64, windows_x64)
2	_listDevices loops over the allocated count instead of the rewritten count.value	reads the count the second call rewrites fails
3	litert_lm_service.dart GPU rejection uses Platform.operatingSystem instead of liteRtLmGateAbi()	3 fails to load with a typed error tests fail
4	N/A (flake fix)	see below

Item 1: new test/fixtures/litert_lm_v1_env_dir_embedding.dart runs the default LiteRtLmRuntimeClient() in a child process with LLAMADART_LITERT_LM_LIB_DIR set to a directory holding the compiled unmarked fixture runtime plus stub companions, with the gate ABI overridden to Linux/Windows; it must refuse with the marker message. A marked runtime through the same route is admitted (control).

Item 4: LiteRtLmBackend.embeddingTimeout is a constructor setting with no per-call override, so the reuse call cannot get its own timeout without a lib/ change. The deadline test now uses a 1 s deadline instead of 100 ms (the assertion is unchanged), which gives the reuse call 10x the headroom. Stability: 20/20 passes under CPU load (3x core count busy loops). The original 100 ms version also passed 20/20 under the same load, so the flake was not reproduced locally; the change is a risk reduction, not a demonstrated fix.

Review Notes

• Independent review status: pending
• CI status / head SHA: pending

High-risk regression review

• Classification: standard
• Implementation task: N/A
• Independent blocking QA task: N/A
• Exact head / current base: N/A
• Production-branch deletion, bypass, or miswire proof: N/A
• Affected-family real-model/artifact evidence: N/A
• Explicit unavailable-family or other N/A evidence: N/A
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

Refs #1837 #1845

[leehack]

Landed through the 0.19.2 batch integration PR #1870 (rebase merge, one commit per constituent; this PR's pinned head was audited/CI-gated there).

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1850.