Track and retire LiteRT-LM downstream speech/runtime workarounds

[leehack]

Summary

Track the downstream compatibility surface introduced by #18 and published in
v0.16.0-native.2, then retire each workaround only after an upstream tagged release and representative real-device validation make it safe.

Current downstream-owned pieces:

• versioned litert_lm_asr_* C ABI over LiteRT-LM's C++-only AsrEngine and AsrSession
• bounded push-PCM AudioSource needed for incremental callers and backpressure
• iOS framework-relative loader paths for the LiteRT Metal accelerator and LiteRT-LM Metal sampler
• checksum-pinned Android Dawn rollback used by the v0.16 runtime on arm64/x64

These should remain explicit compatibility layers, not silently become permanent forks.

Upstream references

• ASR/TTS API exposure: [QUESTION] Question about the new omni folder and planned API exposure google-ai-edge/LiteRT-LM#3058
• Apple framework loader incompatibility: Apple Mach-O bundling: companion dylibs lack -headerpad_max_install_names; gpu_registry/sampler_factory dlopen by basename incompatible with .framework bundling (App Store ITMS-90432) google-ai-edge/LiteRT-LM#2151
• LiteRT accelerator loader PR: gpu_registry: load Metal accelerator from .framework path on Apple google-ai-edge/LiteRT#7219
• Apple Metal sampler packaging/exports: libLiteRtTopKMetalSampler missing from prebuilts / no Bazel target on iOS (Metal backend falls back to CPU sampling) google-ai-edge/LiteRT-LM#1990 and libLiteRtTopKWebGpuSampler exports only 3/7 C ABI functions on macOS/Windows (Linux/Android ship 7) — sampler_factory falls back to CPU sampling google-ai-edge/LiteRT-LM#2073
• broader Mali-G715 runtime tracking: Field report: Gemma 4 E2B + LiteRT-LM 0.10.x on Mali-G715 — 10 findings from sustained Android testing google-ai-edge/LiteRT-LM#2202

Work

ASR C ABI

• Monitor official LiteRT-LM release notes and c/ headers for a supported ASR ABI and packaged ASR symbols.
• Compare any upstream ABI with the behavior downstream currently needs: push PCM, bounded backpressure, confirmed/unconfirmed text, finish/reset, cancellation, and model configuration.
• Prefer migrating to an official ABI over expanding the downstream bridge.
• Keep bridge compatibility until downstream llamadart has passed published-artifact and real-model migration gates.

Apple framework loader patches

• Keep both loader changes declarative and reviewable; replace the imperative sampler_factory.cc text rewrite with a checked-in patch.
• Track the accelerator fix through LiteRT #7219 and obtain an upstream equivalent for the Metal sampler path.
• Remove each patch only when the exact LiteRT/LiteRT-LM tag contains the corresponding fix.
• Revalidate iOS device and Simulator packaging, code signing, framework lookup, Gemma GPU generation, and ASR after removal.

Android Dawn rollback

• File or link an upstream report containing the exact v0.16 Pixel 9 Pro VK_ERROR_DEVICE_LOST reproduction and rollback evidence.
• Keep the override isolated by upstream tag, architecture, source commit, and SHA-256.
• Remove it only after the tagged replacement passes repeated Gemma 4 GPU correctness and stability gates on the affected Pixel/Mali path.

Provenance and compatibility

• Record the resolved upstream Git commit in manifest.json; do not leave upstream.commit null.
• Record every applied source patch and its SHA-256 in the manifest, separately from prebuilt binary overrides.
• Replace the broad v0.16.0 or newer ASR compatibility rule with an explicit tested-version mapping or an equivalent fail-closed capability rule.
• Keep native release tags immutable. Any packaging/provenance correction after .2 must use a new release tag.

Completion criteria

Close this tracker only when every retained workaround has either been removed after the gates above or split into a narrower issue with a named owner, upstream reference, and removal condition.