[leehack]

Fixes #1954

Summary

• Repin WebGPU bridge assets v0.1.83 to v0.1.84 (release 443341300, tag commit 0a3c45f2, bridge source dc359398, manifest SHA-256 55abeb37...04e, native v0.9.0), which adds the thinkingBudgetPauseGrammar capability and the thinkingBudget.pauseGrammar request flag (llama-web-bridge#213).
• WebGpuLlamaBackend sends pauseGrammar: false exactly where native passes pauseGrammarDuringReasoning: false: an eager grammar for which ToolCallGrammarUtils.keepsAppliedThroughThought holds (Qwen3.5 responseFormat, Seed-OSS required calls). Every other request is unchanged on the wire.
• On assets without the capability, that combination throws LlamaUnsupportedException naming thinkingBudgetPauseGrammar and v0.1.84+, instead of returning an unconstrained reply.
• Pin sites updated per native-and-web-sync.md: README, chat app index.html, fetch script, web_bridge_setup.dart, check_webgpu_bridge_tag.dart (release id, tag commit, source commit), tag-check fixtures, docs (doc/webgpu_bridge.md, website page, multimodal embeddings note), changelogs.

Production-readiness scope

• User-facing scope: On WebGPU, responseFormat (and required Seed-OSS calls) with a ThinkingBudget stays constrained after the budget closes the thought, matching native (#1956, #1962).
• Supported platforms/paths: WebGPU/Wasm GGUF via bridge assets v0.1.84+, direct and worker.
• Unsupported or intentionally unavailable paths: Older assets throw LlamaUnsupportedException (names thinkingBudgetPauseGrammar, v0.1.84+) for that combination; other budget requests on older assets work as before.
• 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 (thinkingBudget Dartdoc kept minimal because #1961 also edits it; skills/llamadart-web).
• 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 new public capability flag; validation_coverage_test passes.
• Security/privacy review completed: no secrets, tokens, signed URLs or paths added.
• Follow-up work tracked: None.

PR type guidance

• Bugfix PR: Root cause: the bridge paused every grammar while a budgeted thought was open, so a grammar that takes the template-opened thought resumed inside it and stopped constraining the reply. Fix: bridge flag adopted via the repin, sent under the same condition as native. Regression tests below; affected platform: WebGPU only.

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 N/A as a whole: only tooling/web_setup/structure VM suites are touched; they pass (test/unit/tooling/ test/unit/test_structure test/unit/web_setup)
• dart test -p chrome on test/unit/backends/webgpu (540 tests pass)
• coverage-lib: N/A, changed lib/ code is browser-only (webgpu_backend.dart, interop.dart) and is not in VM coverage; covered by the Chrome suite above
• Other: check_webgpu_bridge_tag.dart --verify-manifest, verify_release_docs_versions.dart, web-lazy-grammar scenario (--list, --dry-run, real run), mutation run below

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyzer, import boundaries	Dart/Flutter 3.47.1	PASS	local
root-vm	tooling, structure, web_setup suites	VM	PASS	local; whole-suite run left to CI
root-chrome	test/unit/backends/webgpu incl. new webgpu_thinking_budget_grammar_test.dart	Chrome	PASS	540 tests
coverage-lib	-	-	N/A	browser-only lib code, see Test Plan
webgpu-bridge-tag-consistency	16 pins agree on v0.1.84; release 443341300 immutable, commit and manifest bytes match	GitHub + CDN	PASS	--verify-manifest
release-doc-version-consistency	install snippets and pins	-	PASS	local
docs-site	website build	-	N/A	docs text edits only; left to CI docs lane
web-lazy-grammar	new thinking_budget_grammar_e2e_test.dart: Qwen3.5-0.8B Q4_K_M responseFormat JSON with budgets 0, 1, 16, direct and worker, CPU Wasm, pinned v0.1.84; control on v0.1.83 asserts unconstrained no-budget JSON and the typed rejection	chrome-webgpu, CPU Wasm	PASS	17/17 incl. lazy and prompt-penalty suites
structured-output-adversarial	native parser/grammar families	-	N/A	no parser, grammar or template change; grammar selection reuses keepsAppliedThroughThought from #1962
high-risk-exact-head-independent-qa	-	-	pending	see below

Mutation check on the new unit test (webgpu_thinking_budget_grammar_test.dart, Chrome):

Variant	Result
origin/main lib (no fix)	2 of 3 fail
Fixed, unmutated	3/3 pass
M1: drop the pauseGrammar: false flag	fails (keeps-applied)
M2: drop the capability rejection	fails (version-skew)
M3: takesForcedOpenThought instead of keepsAppliedThroughThought (misses Seed-OSS)	fails
M4: ignore grammarLazy	fails (stays-paused)
M5: send the flag on every budget	fails

Review Notes

• Independent review status: PASS (fresh auditor-deep at 65c62f4; 0 P1/P2; P3s logged in 0.19.8 audit follow-ups #1950)
• CI status / head SHA: 35 pass, 8 skipped, 0 fail at 65c62f433b.

High-risk regression review

• Classification: high-risk (artifactConsumer, backendRuntime, regressionPolicy)
• Implementation task: lane-1954-webgpu
• Independent blocking QA task: fresh auditor-deep, PASS. It checked the wire contract against bridge dc359398, native parity, the capability probe and the repin (manifest SHA re-downloaded), and simulated 0.19.8 release prep with no drift. 12/12 backend and 10 of 12 tooling mutants killed (T7/T7b shard-weight entries survive, P3). The e2e caught M1 with the WebGPU: thinking budget pauses forced-open-thought grammars (same as #1953) #1954 symptom. Evaluator exit 2.
• Exact head / current base: 65c62f433b / 523ab63a3b
• Production-branch deletion, bypass, or miswire proof: M1-M12 each fail the Chrome unit test; M1 also fails the real-model e2e (direct and worker)
• Affected-family real-model/artifact evidence: Qwen3.5-0.8B Q4_K_M on pinned v0.1.84 assets, direct and worker; v0.1.83 control
• Explicit unavailable-family or other N/A evidence: Seed-OSS has no local weights; covered by the unit test with the Seed-OSS grammar root only. Hardware WebGPU not qualified (CPU Wasm).
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1964.