[leehack]

Summary

• Pin the default WebGPU bridge assets to v0.1.83, built from bridge cb6383fb on upstream llama.cpp v0.9.0@cd959bae. Web now matches the native v0.9.0 pin.
• check_webgpu_bridge_tag.dart: new provenance (release 438242670, tag commit 44aa84a8, manifest ae6f4884…, native anchor v0.9.0); bridgeApprovedNativeUpstreamAhead now equals the bridge's llama.cpp tag, so the native-ahead allowance is closed and the parity wording is required again. The test's approved manifest fixture is the published manifest.json byte for byte.
• Docs drop the Web-behind-native wording; doc/webgpu_bridge.md gains a v0.1.83+ note (Qwen3-TTS suppress tokens, llama-web-bridge#211). The v0.1.82+ capability floors are unchanged.
• Changelogs: one Unreleased bullet each; the native bullet's "Web keeps llama.cpp v0.8.0" sentence is removed.

Release provenance (all first attempts, dispatched from main at cb6383fb, correlation manual-v0.1.83, native v0.9.0 with assets.json SHA-256 0d2119cd…c39e):

Stage	Run	Result
Build Exact Bridge Candidate	42479331139	success
Qualify Exact Bridge Candidate	42480268607	success
Publish Exact Qualified Bridge Assets	42481554696	success

Release 438242670 is immutable; tag v0.1.83 points at 44aa84a8a51701dc9c4f82c21e6ee3e774919c20; the release asset, the tagged manifest.json and the jsDelivr copy all hash to ae6f488408672f9ba28d8f9e664a2b85f4e12048f290e259e18dd0b750c038e5.

Production-readiness scope

• User-facing scope: Web llama.cpp runs upstream v0.9.0, the same release as native; dart run llamadart:web_setup and the chat app fetch v0.1.83.
• Supported platforms/paths: WebGPU bridge direct and worker runtimes (wasm32/memory64), web_setup, chat app bootstrap. Native unchanged.
• Unsupported or intentionally unavailable paths: Unchanged; capability probes and v0.1.82+ floors still reject older assets.
• Out of scope / follow-ups: None.

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.
• 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 added or removed.
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots.
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Feature PR: N/A, runtime pin bump.
• Bugfix PR: N/A.
• Docs-only PR: N/A, changes the pinned runtime artifact.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• dart test -p vm -j 1 --exclude-tags local-only: ran test/unit/tooling/ (1576 pass, 5 skip; 6 windows_registry_qualification_test.dart failures are pwsh missing on this Mac) and test/unit/web_setup (32 pass). Full VM suite left to CI.
• dart test -p chrome --exclude-tags local-only: 4630 pass.
• coverage-lib matrix row: N/A, lib/ changes only two pin constants.
• Other targeted/local-only validation: see Matrix Evidence.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyze, platform boundaries	macOS arm64, Flutter 3.47.1	PASS
webgpu-bridge-tag-consistency	16 pins, Web/native parity, provenance doc, live immutable release and manifest bytes	check_webgpu_bridge_tag.dart --verify-manifest	PASS	release 438242670 at 44aa84a8 with approved manifest bytes
root-chrome	browser unit/integration	Chrome	PASS	4630 pass
examples-tests	chat app test test_vm + Web-only Chrome set	Flutter	PASS	689 + 132 pass
release-doc-version-consistency	install snippets and companion pins	local	PASS	pending companion bumps are pre-existing release-prep items
docs-site	./tool/docs/build_site.sh	local	PASS
chat-app-web-real-model-smoke	built chat app, pinned local v0.1.83 assets, plain chat, penalty probe	Chromium (Playwright 1.61.0), Qwen3.5-0.8B Q4_K_M, CPU	PASS	"2+2 equals 4."; asset URLs ?v=v0.1.83-local-v0.9.0. Playwright 1.64.1 dropped leading keystrokes (prompt typed as "+2? …"), a harness issue; CI pins 1.61.0
chat-app-web-text-to-speech-smoke	chat app Qwen3-TTS synthesis, autoplay, WAV export	Chromium, Qwen3-TTS 1.7B Base Q4_K_M + Q8_0 mmproj, WebGPU offload	PASS	19 frames, 36480 samples @ 24 kHz, 1.52 s WAV
web-speech-task-cancel	real STT/TTS cancellation isolation and reuse, direct and worker	Chrome, Qwen3-ASR 0.6B Q8_0 (jfk.wav exact transcript), Qwen3-TTS 1.7B, wasm64 CPU	PASS	13/13; execution main-thread and worker, core_variant wasm64
web-multimodal-embeddings	Web media embeddings vs native references		N/A	no local EmbeddingGemma-2 native reference fixture; the bridge candidate's durable multimodal gate passed on these exact assets
high-risk-exact-head-independent-qa	independent blocking-only review	Fresh Claude Opus session	PASS	ACCEPT: full audit at d85b6786, delta re-audit at 6caa0736; no blocking findings

Review Notes

• Independent review: ACCEPT (fresh Claude Opus session, uninvolved in implementation). Full audit at d85b6786 (manifest/CDN/release hashes, provenance, mutations M1-M10, Qwen3.5 chat and Qwen3-TTS smokes); delta re-audit at 6caa0736 ACCEPT (only the changelog dedupe differs; chat smoke passes at the exact head).
• CI on 6caa0736: 32 pass, 9 skipped, 0 fail. Head merges main at d53971ee (Speak with Chatterbox and Pocket TTS, with predefined voices #1792); the duplicate native changelog bullet from the union driver was removed by hand.
• Readiness CLI (high_risk_readiness.dart) on the exact head and base: exit 2 (unverifiedPrerequisites, evidence internally consistent).
• Non-blocking follow-ups are tracked on 0.19 small follow-ups (audit P3s, test gaps, doc nits) #1779.

High-risk regression review

• Classification: high-risk (artifactConsumer)
• Implementation task: this PR's author session
• Independent blocking QA task: fresh Claude Opus auditor session, ACCEPT
• Exact head / current base: 6caa07361af44647a815cc268002d425c9bbb4b3 / d53971eefbbae063baea862a4ee61b36d24dfe73
• Production-branch deletion, bypass, or miswire proof: audit mutations M1-M10 on the bridge gate all caught
• Affected-family real-model/artifact evidence: Qwen3.5-0.8B chat, Qwen3-TTS 1.7B (chat app and direct/worker), Qwen3-ASR 0.6B (direct/worker) on the published v0.1.83 assets; see Matrix Evidence.
• Explicit unavailable-family or other N/A evidence: EmbeddingGemma-2 Web media embeddings not rerun locally (no native reference fixture).
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[leehack]

Base drift before merge: main moved from d53971ee to b535dda0, via #1800 (separately audited). The two PRs touch no files in common. git merge-tree is clean, and the combined Unreleased changelog lists each bullet once. The audited head 6caa0736 is unchanged.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1807.