[leehack]

Summary

Tracks #1950 (does not close it). Four small audit leftovers before 0.19.8 release prep:

• Grammar-pause wording (Keep the Seed-OSS required-call grammar applied through a budgeted thought #1962, Sweep #1950 (high-risk): changelog drift gate, generic fallback test, genkit doc floor #1961, Keep response-format grammars applied through a budgeted thought on WebGPU (#1954) #1964 audits): an eager grammar whose root takes the thought (think-body or seed-thought) stays applied through a budgeted thought, whether a responseFormat or required tool-call grammar. Every lazy grammar pauses, whatever its root, and so does every other grammar. Fixed in the thinkingBudget Dartdoc, the generation guide, the WebGPU LlamaUnsupportedException text, skills/llamadart-web/SKILL.md, doc/webgpu_bridge.md, the website WebGPU page and the two Keep response-format grammars applied through a budgeted thought on WebGPU (#1954) #1964 ## Unreleased bullets (reworded, kept in sync, bridge-assets line format unchanged).
• Lazy-grammar test (Pin the lazy-grammar pause and the Seed-OSS call-name constraint in tests #1963 audit): positive check that a lazy grammar still constrains the reply after the thought.
• Drift gate (Sweep #1950 (high-risk): changelog drift gate, generic fallback test, genkit doc floor #1961 audit): docsChangelogDriftProblems names the first differing line and says when only whitespace differs; the docs page must use the prescribed bridge wording (- marker, no "the"), for every retained docs section.
• WebGPU errors (Keep response-format grammars applied through a budgeted thought on WebGPU (#1954) #1964 audit): webGpuBridgeRequestError maps the bridge's CompletionOptions.thinkingBudget and thinkingBudget.pauseGrammar unsupported errors to LlamaUnsupportedException.

Production-readiness scope

• User-facing scope: accurate docs and error text for thinking-budget grammar handling; WebGPU thinking-budget rejections surface as LlamaUnsupportedException.
• Supported platforms/paths: native llama.cpp (test only), WebGPU (error mapping, message), docs and release tooling.
• Unsupported or intentionally unavailable paths: unchanged; assets without thinkingBudgetPauseGrammar still throw LlamaUnsupportedException.
• Out of scope / follow-ups: LiteRT-LM e2e skip, Streaming partial parses differ from the final parse by chunking (MiniMax M3, thinking tags, Gemma 4) #1930, llama-web-bridge items. Shard weights (see below): no change.

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 (existing Keep response-format grammars applied through a budgeted thought on WebGPU (#1954) #1964 bullets reworded; no new bullet).
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant.
• Public capability changes update the validation coverage registry: N/A, no capability change.
• Security/privacy review completed: no secrets or URLs touched.
• Follow-up work tracked: None.

PR type guidance

• Bugfix PR: root causes are loose wording (Keep the Seed-OSS required-call grammar applied through a budgeted thought #1962/Sweep #1950 (high-risk): changelog drift gate, generic fallback test, genkit doc floor #1961/Keep response-format grammars applied through a budgeted thought on WebGPU (#1954) #1964), a lazy-grammar test that passes when lazy grammars are dropped, a normalising drift gate that hides */"the" on the docs page, and an incomplete WebGPU error allowlist (the CompletionOptions.grammar gate excluded thinkingBudget). Regression tests below.

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: touched VM tests (test/integration/thinking_budget_grammar_test.dart with TMPDIR isolated and 2 threads, test/unit/tooling/, all pass); full lane left to CI
• dart test -p chrome --exclude-tags local-only: test/unit/backends/webgpu/ (540 pass); full lane left to CI
• coverage-lib matrix row: N/A, runtime lib/ behavior unchanged apart from an error-mapping branch and message text covered by the Chrome test
• Other: check_platform_boundaries, check_webgpu_bridge_tag.dart --verify-manifest, verify_release_docs_versions.dart plain (pass) and --release-prep (fails only on the expected missing ## 0.19.7 heading and kept ## Unreleased, both changelogs), ./tool/docs/build_site.sh

Mutation results

#	Item	Mutation	Result
2	Lazy-grammar test	never apply a lazy grammar sampler in llama_cpp_service.dart	new test fails; the three older tests still pass (the gap)
3a	Drift message	corrupt the "first differing line" text	killed
3b	Docs-page wording	drop the * marker check / drop the "the default" check	each killed
4	WebGPU errors	drop the thinkingBudget.pauseGrammar alternative / drop the thinkingBudget gate	each killed (Chrome)
1	Wording	docs and Dartdoc only; the ## Unreleased bullets are exempt from the drift gate, so no test can pin them	N/A

Shard weights (#1964 audit T7/T7b): not changed. doc/ci_selection.md makes a missing entry harmless (a new file counts as the median and stale weights only unbalance shards), 29 test files already lack an entry, and the two new entries are placeholders of 0. An "every file has an entry" rule would contradict that contract and fail today; an "every entry exists" rule would only add friction when tests are deleted.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
high-risk-exact-head-independent-qa	independent blocking-only review of exact head	macOS, stories15M, Chrome	PASS	fresh auditor-deep at b9569fb; 17/18 mutants killed (E4 $ anchor predates the PR); 0.19.8 release-prep simulation passes
release-metadata-verification	release docs verifier and drift gate	VM	PASS	verify_release_docs_versions.dart plain and --release-prep as above; test/unit/tooling/ pass
structured-output-adversarial	grammar/structured output	N/A	N/A	no grammar compilation or acceptance change; one test added
litert-lm-v1-public-direct, litert-lm-v1-public-service, litert-lm-v1-public-speech, litert-lm-json-schema, chat-app-litert-json-schema	LiteRT-LM	N/A	N/A	LiteRT-LM untouched

Review Notes

• Independent review status: PASS (fresh auditor-deep at b9569fb, 0 P1/P2; P3s in 0.19.8 audit follow-ups #1950; follow-up bug Granite and Ling 3.0 forced-open grammars pause under a thinking budget #1966)
• CI status / head SHA: see the PR checks; head b9569fba60a0075b210f8934e6f8aad648fd4b96

High-risk regression review

• Classification: high-risk (backendRuntime, releaseAutomation: webgpu_errors.dart, webgpu_backend.dart message, verify_release_docs_versions.dart)
• Implementation task: this lane (worker)
• Independent blocking QA task: fresh auditor-deep, PASS. Simulated 0.19.8 release prep exits 0 and a negative control fails. Evaluator exit 2 at head b9569fba / base c2eb1a82. CI: 32 pass, 9 skipped.
• Exact head / current base: b9569fba60a0075b210f8934e6f8aad648fd4b96 / c2eb1a822d
• Production-branch deletion, bypass, or miswire proof: N/A, no production branch removed; error mapping verified by Chrome test and mutation
• Affected-family real-model/artifact evidence: the lazy-grammar test runs on the stories15M model; no model family behavior changed
• Explicit unavailable-family or other N/A evidence: N/A
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1965.