[leehack]

Summary

Adopts the published llamadart-native v0.9.0 (owner a6046391, llama.cpp v0.9.0@cd959bae, not a prerelease) as the default llama.cpp runtime for core 0.19.0.

• Pins: output of sync_native_release_pins.py --llama-cpp-tag v0.9.0 --litert-lm-tag keep: llamaCppTag, the 11 llamaCppBundleSha256 digests and the windows-x64-cuda12 sidecar digest (all from SHA256SUMS, which check_native_release_digests.dart re-confirms), the Apple XCFramework checksum b0005b1d… in llamaCppCompatibleAppleCompanionChecksums and the companion Package.swift, llamaCppAppleCompanionVersion 0.18.1 (the sync's patch-bump convention, which native_release_pins_test enforces; release prep sets the real number), and the README/installation/support-matrix pins.
• Bindings: sync_native_headers_and_bindings.sh --tag v0.9.0 on Linux arm64 (Docker dart:3.13.1, Ubuntu libclang-dev) reproduces the committed bindings.dart byte for byte, so the earlier hand-generated bindings from native main eb760895 match the published headers.
• Web/native skew: bridge assets v0.1.82 embed llama.cpp v0.8.0, and the bridge gate required Web and native to share an upstream release. Following #1611, bridgeApprovedNativeUpstreamAhead = 'v0.9.0' lets the native pin run ahead on exactly that upstream until a later bridge repin; the gate wording, its tests and both bridge docs say "behind the native runtime". Any other upstream still fails.
• Apple companion: no older runtime is admitted (llamaCppCompatibleAppleCompanionTags stays empty). Earlier frameworks lack the ARM F16 accumulation fix, and published companion 0.18.0 (llamadart: '>=0.18.0 <0.19.0') cannot pair with core 0.19.0 anyway; llamadart_llama_cpp_flutter 0.19.0 ships v0.9.0.
• Hand-updated references (the sync does not own them): pin and Apple checksum tests, runtime_setup_report_io_test, the default-consumer AOT harness (ci.yml input text, native_default_consumer_qualification.yml, verify.py, run_local_e2e.dart, testing_matrix.md and tests now also accept v0.9.0 as a final, non-pipeline tag), the Android validation bindings (tag, android-arm64 archive and headers digests), the Pixel guard tag, the getting-started and embeddings skills, the native-build-hooks override paragraph, and the docs that name the current pin or its grammar/attention behavior (llama-grammar.cpp is unchanged between 8642541b and cd959bae).
• Changelog: one Unreleased bullet in both changelogs: the runtime update, the ARM64 F16 overflow fix (llamadart-native#174), Qwen3-TTS suppress tokens, the companion version and the Web skew.

ABI delta (unchanged from the draft): additive only for this core. mtmd_gen_audio_info gains n_voices and mtmd_helper_gen_audio_inp loses voice; the core references neither, and no symbol it calls is removed.

Production-readiness scope

• User-facing scope: native llama.cpp apps get v0.9.0: large-magnitude F16 dot products on ARM64 CPUs no longer become inf/NaN, and Qwen3-TTS sampling applies the GGUF's suppress tokens.
• Supported platforms/paths: native llama.cpp on every bundled platform; Apple through llamadart_llama_cpp_flutter 0.19.0.
• Unsupported or intentionally unavailable paths: Web stays on llama.cpp v0.8.0 (bridge v0.1.82) until a bridge repin. Chatterbox Turbo/Multilingual and Pocket TTS stay refused by the public adapter with LlamaUnsupportedException (the TTS voices API is #1783). Older Apple companions fail the build as before.
• Out of scope / follow-ups: #1783 (TTS voices, needs #1615); a bridge repin to llama.cpp v0.9.0; packages/llamadart_validation/native/android_qualification/test_prepare.py already fails on main (its fixture still records the v0.8.0-2 archive digest; CI does not run it).

Performance note, not claimed in the changelog: the native release measured CPU F16-weight prefill on Apple M4 at about 0.93x of unpatched upstream v0.9.0 (the accepted cost of the 0006 blocked-GEMM fix); decode and Q8_0 are 0.98-1.00x, and Linux armv8.2_2 F16 prefill is about 1.5x (native#182). No comparison against v0.8.0-3 was measured here.

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.
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant.
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant: bridge gate accepts only the approved ahead upstream (v0.10.0 fails); harness scope tests cover v0.9.0.
• Public capability changes update the validation coverage registry: N/A, no capability change (validation_coverage_test passes in the VM suite).
• Security/privacy review completed: pins, digests and docs only; no secrets or signed URLs.
• Follow-up work is tracked in GitHub Issues, or listed above.

PR type guidance

• Feature PR: runtime update; docs, matrices and changelog updated; negative paths above.
• Bugfix PR: N/A here; the F16 fix's root cause and tests live in native#177/#179/#182.
• Docs-only PR: N/A.

Test Plan

• dart run tool/prepare_workspace.dart: no tracked changes.
• dart format --output=none --set-exit-if-changed .
• dart analyze: no issues (also packages/llamadart_validation).
• dart run tool/testing/check_platform_boundaries.dart
• TMPDIR="$(realpath "$(mktemp -d)")" dart test -p vm -j 1 --exclude-tags local-only: 11491 passed, 129 skipped, 6 failed, all 6 in windows_registry_qualification_test.dart (no pwsh on this host; same on main).
• dart test -p chrome --exclude-tags local-only: N/A locally, no Web-reachable code changed (bridge gate is tooling); CI runs it.
• coverage-lib: N/A, lib/ changes are pin constants and a doc comment.
• Other: check_webgpu_bridge_tag.dart, check_native_release_digests.dart, the four native-sync Python suites (Python 3.15), the default-consumer harness Python suites, verify_release_docs_versions.dart, validation package tests (641) and Android qualification Python tests, and real-runtime rows below against the published v0.9.0 through the normal hook (bundle 7ec293c7… downloaded into native_bundles_v2/v0.9.0, no path override).

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
root-vm	full VM suite	macOS arm64, published v0.9.0	PASS except 6 env failures	windows_registry_qualification_test needs pwsh
native-release-digests	bundle, SD and LiteRT-LM digests vs releases	network	PASS	llamaCppBundleSha256 matches v0.9.0 SHA256SUMS (12 bundles)
native-sync-python-suites	pin policy, header archive, bindings sync, SD bindings	Python 3.15	PASS	all four OK
native-hook-bundles	hook, bundle selection, companions	macOS arm64	PASS	part of root-vm
text-to-speech-smoke	Qwen3-TTS capability, synthesis, seed determinism, cancel	Qwen3-TTS-12Hz-1.7B Q4_K_M + mmproj Q8_0, CPU	PASS (6/6)	103680 samples @ 24 kHz
validation-speech (tts)	lifecycle, cancel, reload, bounds	locked pack, CPU	PASS (25/25)	RTF 0.77-1.01, cancel 13.8 ms
validation-speech (stt)	Qwen3-ASR WER, edge fixtures, cancel	locked pack, CPU	PASS (28/28)	WER 0.0 on the reference clips
gguf-chat-features-smoke	chat, thinking budget, tool calls, image input	Qwen3.5-0.8B Q4_K_M + mmproj F16, CPU	PASS	all 8 cases finish as expected
native-exit-teardown	C exit idle/generating/loading/projector/disposing, errors, kill	Qwen3.5-0.8B + mmproj, macOS Metal	PASS (9, 1 skipped)	decision-head variant not configured
release-metadata-verification	release docs pins	local	Expected pending	default mode passes with pending companion bumps; --release-prep fails only on the version bumps and Unreleased headings that 0.19.0 release prep owns (same LiteRT-LM items as main)
structured-output-adversarial	grammar limits	N/A	N/A	only a doc comment and docs name the new upstream; src/llama-grammar.cpp is identical between 8642541b and cd959bae; grammar suites pass in root-vm
litert-lm-v1-public-*, litert-lm-json-schema	LiteRT-LM	N/A	N/A	LiteRT-LM pin unchanged
high-risk-exact-head-independent-qa	exact-head audit	see below	Pending	requested on the current head

Review Notes

• Independent review status: pending; the operator runs the exact-head audit.
• CI status / head SHA: see checks on the current head.

High-risk regression review

• Classification: high-risk (artifactConsumer, backendRuntime, regressionPolicy, structuredOutput)
• Implementation task: this PR
• Independent blocking QA task: ACCEPT from a fresh Claude Opus auditor session (operator-owned, no part in the implementation): no blocking findings. Main unmoved, merge-tree clean; CI 38 pass / 8 skipped / 0 fail; native digests, bindings (byte-identical to a fresh header sync), ABI (11 externs added, none removed), bridge-gate and AOT-scope mutations all independently verified. Maintainer approved merging on ACCEPT and kept the Web-bridge ahead allowance.
• Exact head / current base: bc724a2b0c71c8aaaf2f31edb30b23316dd4671f / af56657fa061b5c7621ecb5517199f6683ff10b1. Readiness CLI (tool/testing/high_risk_readiness.dart) exits 2 (unverifiedPrerequisites, internally consistent) with evidence citing the pin, bridge-gate, workflow, AOT, Android-profile, grammar and bindings tests, each with a head pass and a failing control.
• Production-branch deletion, bypass, or miswire proof: the bridge gate's ahead allowance is exact-match (v0.10.0 is refused in a test); the normal hook downloaded and verified the published v0.9.0 macOS arm64 bundle by digest, with no path override.
• Affected-family real-model/artifact evidence: Qwen3-TTS, Qwen3-ASR and Qwen3.5 VL on the published artifact (Matrix Evidence).
• Explicit unavailable-family or other N/A evidence: Windows, Linux, Android and iOS hosts not run locally; Apple companion SwiftPM build not run locally; no F16-weight model run (the native release qualified the F16 kernel).
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1784.