[leehack]

Summary

• Adds chat-litert-auto, the chat-litert-gpu profile under backend: auto, so the Galaxy S24 ComputeDevice.auto GPU-to-CPU retry can run from a clean committed bundle. Remote runs now refuse dirty bundles, so the local profile copy in the #1856 recipe no longer works.
• Records the paid Galaxy S24 Firebase run for #1458 item 4 in doc/cross_platform_validation.md. Result: GPU refused, one warning per engine, CPU retry, 16/17 PASS.
• The harness auto mapping that item 4 asked for was already fixed in #1658. I re-checked it on d739cfdb8: the test LiteRT profiles preserve automatic and explicit device selection passes, and an auto -> GpuBackend.cpu mutant fails it (Expected: GpuBackend.auto, Actual: GpuBackend.cpu).

Production-readiness scope

• User-facing scope: None. This changes the private validation harness and maintainer docs only.
• Supported platforms/paths: Android validation through Firebase Test Lab. The profile is also bundled into the QA app's profile list.
• Unsupported or intentionally unavailable paths: Unchanged.
• Out of scope / follow-ups: #1903: after a cancel, LiteRT-LM 0.19.1-2 exposes no performance counters, so C08.cancel is NOT_RUN. It also happens on macOS CPU, so the retry does not cause it.

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly with actionable diagnostics or are clearly documented as unavailable. (Unchanged.)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (Validation README and runbook updated. No changelog entry: the private harness is not shipped to users.)
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant. (The profile config hash is pinned in embedding_audit_test.dart; the auto mapping test exists from fix: preserve automatic LiteRT backend on context recreation #1658.)
• Public capability changes update the validation coverage registry with a pack or a reasoned, issue-linked exemption; inventory checks pass (or N/A with scope). N/A: no capability change.
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots. (The docs quote only the matrix IDs and the model hash.)
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above. (LiteRT-LM 0.19.1-2: no performance counters after a cancelled generation (C08.cancel NOT_RUN) #1903)

PR type guidance

• Feature PR: N/A, no capability.
• Bugfix PR: N/A, no defect fixed. The auto mapping fix is already on main (fix: preserve automatic LiteRT backend on context recreation #1658).
• Docs-only PR: Docs, plus one validation profile and its pinned hash. No runtime behavior changes.

Test Plan

• dart run tool/prepare_workspace.dart (clean on d739cfdb8)
• dart format --output=none --set-exit-if-changed . (0 changed)
• dart analyze: N/A at the root, since only the package changed. packages/llamadart_validation dart analyze: no issues.
• dart test -p vm -j 1 --exclude-tags local-only: Markdown-selected root suites (tool/ci/markdown_test_suites.txt) plus litert_lm_tokenizer_fixture_test.dart pass, 815 passed and 4 skipped. The full suite runs in CI.
• dart test -p chrome --exclude-tags local-only: N/A, no web or lib/ change.
• coverage-lib matrix row: N/A, no lib/ change.
• Other: packages/llamadart_validation dart test -p vm: 642 passed. example/chat_app flutter test test/validation_app_test.dart: 3 passed (bundled profile list). Firebase device run below.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyzer	macOS arm64, Flutter 3.47.1	PASS	0 files reformatted; package analyze clean
validation-harness	private runner, profiles	macOS arm64, model-free	PASS	642 passed; new profile hash pinned
root-vm	Markdown-selected tooling suites	macOS arm64	PASS	815 passed, 4 skipped
root-chrome	browser lane	N/A	N/A	no web code
coverage-lib	lib coverage	N/A	N/A	no lib/ change
Firebase Android device run	ComputeDevice.auto GPU failure and CPU retry	Galaxy S24 SC-51E API 36, Adreno 750; Qwen3 0.6B legacy LiteRT artifact; LiteRT-LM 0.19.1-2; source 3f6338b50 (clean)	16 PASS, 1 NOT_RUN (#1903)	matrix-1da3zj25gmeu1; details below
high-risk-exact-head-independent-qa	independent review	N/A	N/A	classifier: standard

Review Notes

• Device run (matrix-1da3zj25gmeu1, test time 213 s):
  • Every engine creation tried the GPU first: the first use, C09.reload and C12.recovery.
  • The runtime refused each attempt with FAILED_PRECONDITION (9): Android GPU safety restriction: observed Vulkan device vendor=20803 ... maxStorageBufferRange=134217728 ... 155582464-byte constant ... Use CPU.
  • llamadart logged [WARN] LiteRT-LM could not create a GPU engine under ComputeDevice.auto; retrying on the CPU. Load with ComputeDevice.cpu to skip the GPU, or ComputeDevice.gpu to fail instead. once per engine (3 in total), then XNNPACK created a CPU engine.
  • The warning came 0.48 s, 0.91 s and 0.62 s after each load.
  • After the retry, diagnostics report LiteRT-LM cpu with 0 GPU layers. C01.load still reports gpu because engine creation is deferred.
  • About 8 tok/s wall for 32 tokens (B01). No crash and no process death in logcat. Collection complete, cleanup verified.
• First attempt (matrix-1tcmnx5y9bnqv, test time 602 s) stopped at the harness's 10-minute model download deadline at 260777210 of 614236160 bytes, before any case ran. The download ran at about 3 MB/s, then slowed to about 100 KB/s. I retried once.
• Cost: about 15 billable physical device-minutes (11 + 4, rounded up per run). That is about USD 1.25 gross at USD 5/device-hour, before the daily free allowance. Personal project only, Blaze guard cap USD 2.50.
• LiteRT-LM 0.19.1-2: no performance counters after a cancelled generation (C08.cancel NOT_RUN) #1903 attribution: a local macOS chat-litert-cpu run at d739cfdb8 has the same C08.cancel NOT_RUN with null counters. At 0.18.1-4, three Linux CPU runs passed with counters.
• Independent review status: N/A, standard risk.
• CI status / head SHA: pending, head 7584d2e141bc33b13ea25043cad3f646fbd151e5.

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

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1904.