Make zero-temperature LiteRT generation greedy

[leehack]

Summary

Native LiteRT-LM requests with temp: 0 and the default topK: 40 can produce corrupt Windows GPU output. Request a single greedy candidate for zero-temperature raw generation and chat; preserve positive-temperature sampling settings.

Production-readiness scope

• User-facing scope: Correct greedy behavior for zero-temperature native LiteRT-LM generation through llamadart.
• Supported paths: Native service raw generation and native chat. Existing platform availability is unchanged.
• Unavailable paths: Stock Linux/Windows GPU selection remains gated until the separately qualified runtime is published and adopted. This PR does not change the native pin or enable a platform.
• Follow-ups: Native runtime packaging is tracked by Fix shared-runtime linkage for desktop GPU builds litert-lm-native#50 and the desktop investigation by Diagnose NVIDIA L4 desktop WebGPU initialization crashes before enabling consumer GPU litert-lm-native#47. The low-level upstream sampler itself is not patched.

Validation

• Pinned SDK workspace preparation, repository formatting and analysis passed.
• All 61 native service tests passed. Six new cases cover raw/chat zero-temperature requests with top-k 0, 1 and 40; existing positive-temperature cases verify unchanged forwarding.
• Removing the normalization causes the raw/chat regression cases to fail.
• Full VM suite passed: 2,085 tests, 77 skipped; line coverage 78.62%. Chrome passed 925 tests. Docs build and link validation passed. All 15 final-head GitHub checks passed.

Matrix row	Platform/model/backend	Result	Evidence
windows-x64-ci-runtime (targeted device extension)	Windows Server 2022, NVIDIA L4, driver 582.53, D3D12; Qwen3 0.6B and Gemma 4 E2B	PASS patched service	Fresh math/capital answers, streaming cancellation and reuse. Unpatched Qwen service fails with the same candidate/model/driver. Qualification-only service subclass exposes the known test GPU; stock platform gates stay intact.
windows-x64-ci-runtime	Same Windows VM, repaired Qwen3 CPU	PASS	Patched-service CPU regression control.
macos-arm64-runtime-smoke	Apple Silicon, repaired Qwen3, published v0.17.0-3 Metal	PASS	Existing public-engine chat/thinking/history smoke; strict tool-choice rejection retained.
web-chrome-runtime-smoke	LiteRT browser regression	PASS	925 Chrome tests; Web sampling implementation is unchanged.

Windows native inputs are the final PR50 candidate at 4c66cb9f1f037f91eed9b1cbe9101463efd8e46a, with upstream-pinned DXC 1.9.2602. Qwen3 bundle revision 8414150f2e9dcc82449bcc9c5abc404b399a4d06 uses the original-tokenizer repair merged in native PR49. Gemma bundle revision b3ca0d2f076785a8f4b2219ddbd2bdb99954eae1 uses gemma-4-E2B-it.litertlm.

Additional isolation: both models fail GPU sampling with temperature 0/top-k 40, pass top-k 1, and pass temperature 0.8/top-k 40. Removing the optional sampler in a test copy and using an explicit library path confirms native CPU sampling fallback while D3D12 model execution remains active; both models pass that control. The temporary VM, disk and firewall rule were deleted after evidence collection.

High-risk regression review

• Classification: high-risk (backendRuntime)
• Implementation task: Windows sampler isolation and greedy-service fix in this thread.
• Independent blocking QA task: /root/litert_greedy_adversarial_audit: complete, no blocking findings; eight independently rerun focused tests passed.
• Exact head / current base: f6132610181734d2411ed7161b789adc8e6ca9c5 / a5df1c4fcbb1766d26efb5b1d9becda191df89a3.
• Production-branch deletion proof: Removing both normalization expressions fails raw and chat regression tests for zero-temperature multi-candidate requests.
• Affected-family real-model evidence: Windows Qwen3/Gemma4 service before/after controls; Mac public-engine Qwen3 Metal smoke.
• Unavailable-family/platform evidence: Pixel and iOS were not rerun for this change; platform gates and runtime pins remain unchanged. Web real-model GPU was not exercised.
• Known PR-caused P1 regressions: 0 known after independent review and completed validation.
• Unresolved review threads: 0 independently verified at final head.

The repository-local readiness evaluator reports internally consistent evidence with unverifiedPrerequisites: external authenticated enforcement prerequisites are unavailable. This is not a claim that those prerequisites are configured. The independent exact-head review above is complete.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-508.