Qualify post-v0.4.0 upstream grammar and GDN correctness fixes

[leehack]

Known upstream correctness fixes missing from published v0.4.0

Current native v0.4.0 and Web v0.1.43 bundle upstream 5266f24da75dc449bd56cbed7addb9c8e4a6a73e. They do not contain:

• GDN normalization fix models: fix GDN normalization from max to rsqrt ggml-org/llama.cpp#28068 (5fdfa6282936576d2f352d4b97f397a109f207a6): Qwen3.5/MoE/Next/4exp, Kimi Linear/K3, BailingMoE3.
• Grammar repetition threshold fix grammar : fix max repetition threshold ggml-org/llama.cpp#28469 (9e0e220594af405a62835dc3a27495729fd8506b): exact threshold 2000 is erroneously rejected.
• Metal command-queue failure leak metal : fix memory leak in early return ggml-org/llama.cpp#28399.
• CUDA matrix races cuda: fixes races in mmid and mmf ggml-org/llama.cpp#28475 and divergent barrier ggml-cuda: fix divergent barrier in f16 flash attention ggml-org/llama.cpp#27870. NVIDIA hardware is unavailable; do not claim sanitizer clearance.

These are upstream defects, not evidence of a Dart integration regression caused by PR #478. Published artifacts remain immutable.

Qualification candidate is exact upstream main 73ab7599b553c03f6f5d2db24a18ad76f2eb36a3 (31 commits beyond v0.4.0). Qualification is not permission to publish a nightly or change stable-channel policy.

Acceptance: reproduce applicable public grammar boundary; run candidate upstream grammar tests and affected-model A/B checks; review ABI/header changes; qualify owning native wrapper and matching Web build/model matrix before downstream pins. Keep native/Web source ownership separate and zero-patch in llamadart. Record hardware/model omissions explicitly. Select an upstream-aligned release identity through existing policy; do not invent a stable v0.4.0 rebuild label for different upstream source. Publication and coordinated downstream adoption require their separate approval boundaries.

[leehack]

Qualification update (2026-09-07):

• Current public Dart path reproduced the grammar defect with the published native v0.4.0 runtime and stories15M GGUF. Generation with root ::= "a"{1999} succeeded; root ::= "a"{2000} threw LlamaException: llama.cpp failed to initialize the requested grammar sampler. Reproduction uses LlamaEngine.loadModel/generate, not only a parser helper.
• Exact upstream candidate 73ab7599b553c03f6f5d2db24a18ad76f2eb36a3 built locally with CPU/Metal. Both test-grammar-parser and test-grammar-integration passed (2/2); both sources explicitly undefine NDEBUG, so Release assertions remain active.
• Candidate llama-cli with the same stories15M model and exact {2000} grammar generated a and exited successfully. This establishes the corrected upstream boundary, not yet a published owner/consumer artifact fix.
• Local evidence: /private/tmp/grammar-boundary-v040.log, /private/tmp/post-v040-grammar-tests.log, /private/tmp/post-v040-grammar-cli.log. Qualification checkout is isolated; no upstream patch, release pin, publication or workflow dispatch occurred.

Still required: affected GDN model comparison and owner wrapper/Web qualification; approved release identity. Current latest stable remains v0.4.0. Do not relabel this later upstream commit as a v0.4.0 wrapper-only rebuild.

[leehack]

Additional read-only/local qualification evidence

Compared published v0.4.0 upstream 5266f24da75dc449bd56cbed7addb9c8e4a6a73e with exact candidate 73ab7599b553c03f6f5d2db24a18ad76f2eb36a3.

• Same checksum-pinned Qwen3.5-0.8B Q4_K_M (bd258782e35f7f458f8aced1adc053e6e92e89bc735ba3be89d38a06121dc517), macOS Metal, seed 1/temp 0, context 2048, 96 tokens, reasoning off: both builds completed and produced the same final sky-color explanation. This is one bounded integration smoke, not a quality/performance benchmark or coverage of all GDN families.
• No diff in the compared public llama.h, mtmd.h, mtmd-helper.h, and ggml.h headers. This alone is not full ABI qualification.
• Unmodified bridge source 89178be67c3c84300bc1b129182bd5bc5a8e21fc built both wasm32 and memory64 against the exact candidate with Emscripten 6.0.8. Model-backed state persistence and Qwen3.5 image inference passed direct and worker modes. These local browser smokes do not claim hardware WebGPU qualification.
• Candidate ASR/TTS, full native owner/platform matrix, and CUDA hardware remain unqualified. Pixel testing is separately targeting the published v0.4.0 artifact, not this candidate.

Publication boundary remains: do not disguise changed upstream source as a v0.4.0 wrapper-only rebuild. A corrected stable artifact needs a containing upstream stable release or a separately decided development-channel policy. No pin, tag, release, or workflow dispatch was performed.

[leehack]

Local qualification follow-through on exact candidate 73ab7599b553c03f6f5d2db24a18ad76f2eb36a3 is now complete for the available bounded gates:

• Unmodified bridge 89178be67c3c84300bc1b129182bd5bc5a8e21fc, Emscripten 6.0.8, existing wasm32/memory64 candidate: checksum-pinned Qwen3-ASR cold/warm transcript, cancellation, silence passed all four memory/runtime modes; Qwen3-TTS memory64 direct/worker synthesis/cancel/pre-abort/reuse/unload passed. TTS logged Apple metal-3 WebGPU active; ASR harness deliberately uses CPU layers. Speaker-reference TTS was not exercised.
• Current native owner 6a30a10908aa01c8c15b8d1fb5fd016cd9e0c721 with only the isolated upstream candidate substitution compiled macOS CPU+Metal Release. CTest 3/3 passed with active assertions; required exports 36/36 passed.
• No new candidate regression was found. This is local evidence, not hosted release qualification, a full platform matrix, CUDA clearance, or published artifact adoption. Pixel separately qualifies published v0.4.0.

Preparing a nonpublishing CI-only candidate qualification PR to use existing PR-triggered Windows ARM64, Android ISA, emulated dispatch, Linux wrapper/artifact and Windows x64 link checks. Production submodule and stable-only release policy stay unchanged. Latest upstream stable remains v0.4.0; no later source may be mislabeled as its wrapper rebuild. No workflow dispatch, tag, publication, ingestion or downstream pin change occurred.

[leehack]

Qualification is now running through nonpublishing draft PR #77 at exact head 2d3bdce4cc64cf97bf9f4b0358c6f631adf8a66f, without a manual workflow dispatch or production pin change.

The first hosted run passed 13/14 checks (candidate and pinned Windows ARM64, Linux wrapper/artifacts, Windows x64 link, emulated KleidiAI and provenance). Android candidate compilation passed, then the existing source audit correctly rejected the previously unknown candidate ggml digest. Independent audit confirmed CPU feature detection, KleidiAI selectors/tables/callers and CPU build inputs were unchanged. After explicit maintainer approval, PR #77 added only the exact reviewed ggml/Kai pair while preserving v0.4.0's pair and every binary-containment restriction. Unknown/cross-combined pairs and new callers remain rejected.

Final-head independent Astra QA passed, local focused ISA 19/19 and full Python 119/119 passed. Replacement full 14-check hosted matrix is running; especially Android disassembly must pass before calling qualification ready. This remains preparation, not publication or full physical-device/accelerator qualification.

[leehack]

The approved exact-source pair is now fully checked by replacement PR CI. Draft PR #77 head 2d3bdce4cc64cf97bf9f4b0358c6f631adf8a66f has 14/14 successful checks; final independent Astra QA passed; zero review threads/unresolved findings.

• Android candidate audit: 256,191 instructions inspected; 2,187 scalable instructions confined to 18 exact approved ELF functions.
• Pinned and candidate Windows ARM64 build/CTest, Linux x64/arm64 artifact+clean dlopen, Linux wrapper, Windows x64 link all passed; candidate public grammar sampler accepts exact2000 and rejects2001/malformed grammar.
• Emulated KleidiAI dispatch and compute passed. Source fingerprints remain paired; all existing binary restrictions remain intact.
• Runs: https://github.com/leehack/llamadart-native/actions/runs/34167970481 and https://github.com/leehack/llamadart-native/actions/runs/34167970483 .

PR remains draft, unmerged. This completes the bounded nonpublishing owner CI slice, not #76's artifact publication/adoption acceptance. Full release artifact matrix, hosted exact-Web-artifact attestation and appropriate upstream release identity remain separate. CUDA/older Android hardware remains unavailable/unverified; Pixel evidence is for published v0.4.0, not candidate73ab7599. No manual dispatch, release, tag, publication, ingestion or production/downstream pin change occurred.