test/multiple-subtitles-ass.test.js verifies hashes of actual core5.1.2 main, legacy and source before comparing conversion behavior. It covers the collapsed main output, source-equivalent timings/text, valid/custom/empty passthrough and converter failure cleanup through the real candidate plugin. It runs in both test:unit and test:multiple-subtitles. This is native VTT conversion, not ASS layout/rendering acceptance for JASSUB.

test/multiple-subtitles-caption.test.js checks the old-host view adapter with the actual plugin bundle: native order, original update identity/scalar event payload, escape changes, timestamp cleanup, modern-host isolation and destroy/registration failure cleanup. It runs in test:unit and test:multiple-subtitles, without new dependencies. Native cue timing is separately verified in the browser suite.

test/multiple-subtitles-entities.test.js checks named entities, literal tags, double encoding, nested semantic tags and selection/reset through the real plugin. The vendor parity suite stays unchanged: the plugin uses the parser's existing entity-table option. Historical malformed-tag and stray-semicolon tests continue to assert old defects; corrected cases have explicit rendering expectations instead of treating those old serialized bytes as a required contract. Browser entity tests compare native VTT fragments and actual HTML captions on both core variants.

yarn test:library runs actual library builds, build-input freshness checks and the scaffold's generated consumers. The new build modules live in scripts/library/; yarn typecheck:library checks them and the retained JS/MJS entrypoints. Browser development/watch/error-recovery coverage is browser/library-development.spec.js.

yarn test:dev-server covers actual HTTP ranges/HEAD/gzip, root containment, occupied-port CLI failure without stopping the existing server, repeated same-port reuse with SSE connections, aborted downloads and startup/build cancellation. A controlled-clock test checks heartbeat cleanup separately from real sockets. It is included in test:node and test:library; the browser suite also verifies actual docs/Monaco TypeScript execution and local media playback through this server.

yarn test:scaffold covers old plugin-generator defects, exclusive writes and rollback, the historical CLI path, and actual generated package builds and type consumers. It runs in ignored fixtures without adding a workspace or real demo; see scripts/plugin/README.md. yarn typecheck:scaffold checks its TS modules and JS command shim. Both are wired into the existing Node/CI checks.

site-loading.test.js freezes the old mobile loader's failure/race reproduction and checks shared loader ownership, retries/cancellation, query encoding and language rules. Actual local pages and Monaco are covered by site-loading.spec.js.

Editor tooling lives in scripts/editor-declarations/; see its README. test/editor-types.test.js checks standalone globals and module consumers with current/old compilers, Chapter/VAST regressions, generated files and library-list failures. Package baseline tests continue to exercise the old MJS import paths.

Use the pinned Node/Yarn toolchain from ../refactor/toolchain-setup.md.

yarn build:test generates deterministic documentation readiness smoke and its source manifest. yarn check:docs-smoke is read-only, and yarn typecheck:docs-tools checks the TS modules and legacy JS command shim. refactor/scripts/docs-smoke.test.mjs covers the historical malformed-parser loop, all 233 current source snippets, stable generation and CLI drift/error handling. test/browser/docs-smoke.spec.js uses actual core/media in three engines for readiness, failures, frame cleanup, storage restoration and repeated execution. See scripts/docs-smoke/README.md for the explicit limits of readiness smoke.

yarn test:vue-consumer checks the actual Vue example with packed core/Danmuku/ Document PiP in an outside-workspace install. It also installs its pinned Vue compiler there, checks all declaration paths for isolation, and runs development and production builds in three engines. test/vue/ contains strict SFC/type fixtures and a separate plain-JS consumer. Reports under refactor/.cache/vue-consumer-* cover updates, remount, KeepAlive, captured errors, native media and real Worker cleanup. --before substitutes the recorded old wrapper and must preserve the same lifecycle behavior; it is not expected to fail.

yarn test:react-consumer installs packed core/Danmuku/Document PiP outside the workspace and checks the actual React example's strict TSX, development and production builds, and native lifecycle/playback in three engines. Its fixture is in test/react/; unique reports and frozen consumer locks are retained under refactor/.cache/react-consumer-*. The --before control intentionally fails at the recorded old wrapper's callback-exception leak. See the example README for the compatibility contract and the capabilities covered by other package tasks.

test/auto-thumbnail-encoding.test.js verifies a separate30-second JPEG deadline, withheld and duplicate callbacks, replacement/destruction, synchronous completion and timer cleanup failures. It runs in test:auto-thumbnail and test:unit. Native lifecycle tests deliberately hold an actual JPEG callback and invoke the captured deadline; that proves cleanup with native resources, not a spontaneous encoder hang. First-frame pixel and device limitations remain separate gates.

yarn test:auto-thumbnail-types checks the preserved npm 1.1.0 declaration, accurate /runtime Promise types, exact rejected consumer lines, and generated editor globals. It also runs in test:auto-thumbnail and test:baseline. yarn test:auto-thumbnail-types-package packs the candidate and installs it and the actual npm 1.1.0 archive outside the workspace with a packed core. It verifies frozen reinstalls, byte identity, conditional declaration resolution, old/current compiler consumers, no-interop CommonJS, and installed factory registration. It also installs actual 1.0.1 to verify old .default calls and confirms 1.0.0's missing main/legacy artifacts. Candidate direct/default calls share a callable; runtime types describe the alias without changing the historical root declaration. This does not decode video or close all older type-shape, native pixel, or device acceptance gaps. Candidate entry types use the same runtime files.

yarn probe:auto-thumbnail-rendering is an optional diagnosis command, separate from passing test suites. It runs four intrinsic-size rendering modes and three readiness strategies in Windows WebKit, plus Chromium/Firefox controls. It records raw draw pixels, seek/event ordering, frame counters and exact code/media hashes in a new cache directory. Completion of the command does not imply correct pixels: the current WebKit first-frame discrepancy remains an open release/refactor risk.

yarn test:danmuku-mask runs candidate run cancellation/resource ownership plus frozen historical defects and public contracts. Candidate tests also run in test:unit; their controlled SDK, RAF and canvas hosts do not prove native model inference, GPU disposal or browser mask geometry. See the Mask architecture map. yarn test:danmuku-mask-types-package installs both published versions and the candidate tarball outside the workspace for old/current compiler compatibility, negative cases and CJS/legacy registration without model startup. The ordinary baseline suite also checks public declaration identity and editor generation.

yarn test:mediabunny runs historical lifecycle observations, candidate load cancellation, and real SDK input parsing/track contracts; all are in test:unit. Candidate failures can be reproduced against the frozen main with ARTPLAYER_MB_BASELINE=1. This does not replace native playback or installed consumers. See refactor/mb-validation.md and the proxy package's ARCHITECTURE.md for browser and artifact selection.

yarn test:ads runs source, verified npm 1.0.6 and the frozen unpublished 2.1.0 Ads bundle against controlled clock/host contracts. It is included in test:unit. The helper uses the real option validator but does not simulate media decoding or layout. Historical defect assertions apply only to frozen implementations. See Ads validation for artifact overrides and browser scope. ads-lifecycle.test.js adds candidate-only resource, reentry, early-method and Promise regressions. Async observations drain one event-loop turn, including VM Promise assimilation, rather than assuming a fixed number of microtasks. The published defect observations stay intact.

Command	Scope
`yarn test:unit`	Playback and DASH regressions, public contracts against released/current code, and JS/TS fixture loading
`yarn test:node`	Unit contracts plus toolchain and documentation build failure propagation
`yarn test:ci`	Actual CI summary exit codes, workflow regression guards and repository impact analysis
`yarn test`	Node checks and the committed baseline/tooling tests in refactor/scripts
`yarn ci:check`	Toolchain, plan, read-only lint, types, then yarn test
`yarn test:imports`	Existing distribution import smoke examples; run after building

helpers/load.js owns source and artifact selection. loadModules bundles named internal modules with esbuild and requires exactly one JS/TS source file. loadPackage uses the repository Vite configuration without writing distribution files; it requires a single self-contained JavaScript chunk. These loaders transpile but do not replace yarn typecheck.

helpers/playback.js creates the controlled media facade used by the original playback regressions. It models only the exercised methods and events. It is not HTMLMediaElement validation; actual playback remains in browser tests.

node --test test/hls-control.test.js runs shared HLS control contracts against workspace source and integrity-checked published main/legacy/ESM. It is included in yarn test:unit. helpers/hls-control.js owns the controlled SDK/registry host; it does not simulate decoding or ABR. Published-only tests preserve historical bug observations without requiring candidate code to keep them. Real Hls.js 1.5.17 playback is in test/browser/hls-control.spec.js; see its environment limits in refactor/hls-validation.md. Test SDK download is pinned and verified, not an installed runtime dependency.

contracts/emitter.js owns public assertions, independent of module layout. public-behavior.test.js runs them against the integrity-checked published core and current bundled source. The combined chain/context contract corresponds to BASE-03 EVENT.chain and EVENT.context-arguments; other EVENT IDs retain their baseline names. Exceptions, callback identity, mutation during dispatch and once reentry must not be weakened during migration.

Set ARTPLAYER_TEST_CORE to a built .js, .legacy.js or .mjs core file to add a candidate to the same public-contract run, then execute node --test test/public-behavior.test.js. An invalid path/export fails the run. This is a narrow event contract check, not isolated package installation or complete API compatibility; ENG-07 owns tarball consumption.

When a production module moves, update its loader mapping and its maintenance documentation, preserving the behavioral assertions. New contracts belong in contracts/; test-specific controlled state belongs in helpers/. Do not modify frozen refactor/fixtures or baseline captures just to pass changed behavior. Record historical defects and candidate fixes separately.

ci-summary.test.js exercises the real CLI with success, cancelled and malformed provider data, plus missing groups and metadata/environment files. It is included in test:node. refactor/scripts/ci-workflow.test.mjs parses the actual YAML and rejects broken matrix, cache, install, report and summary requirements; it is included in test:baseline. These local tests do not replace hosted runner or branch-protection evidence. See CI operations before changing a required job or its cache.

Real browser tests

test/utils.test.js compares the published and current utilities, including exact formatting, subtitle text, property/merge behavior, controlled timers and the full utility export surface. It accepts the same ARTPLAYER_TEST_CORE option for all three artifact formats. The merge prototype fix is a candidate assertion alongside a published-only defect observation. test/types/utils-source.ts checks strict internal source inference without changing the still-separate public declaration contract. Browser utility tests exercise actual downloads, Blob URL contents and cleanup when a controlled click fails.

See browser/README.md for yarn test:browser, browser installation, published/current combinations, media fixtures, candidate mapping and failure reports. These tests are separate from the fast Node suite and run in the Browser playback smoke CI job.

test/audio-track.test.js compares source and published Audio Track 1.1.0 in all three formats with a controlled Audio object: timing intent, offsets, source updates, rejection, volume/rate, independent instances and cleanup. Published-only tests retain defect evidence. See audio validation for actual media tests and limitations.

Document PiP: yarn test:dpip runs 48 historical window/DOM/lifecycle cases. Native DOM iframe checks use yarn test:browser test/browser/dpip.spec.js; controlled window APIs do not establish native Document PiP support. See validation notes.

Danmuku: yarn test:danmuku covers historical behavior and current input, scheduling, settings, rendering and heatmap boundaries. ARTPLAYER_DANMUKU_ARTIFACT selects a built main/legacy candidate for integration cases; direct internal tests remain source tests. This controlled synchronous VM loader executes CommonJS/UMD scripts, not .mjs modules. Do not pass ESM bytes to that override or interpret its parser failure as a production ESM failure. Actual ESM imports are checked by the isolated package consumer; full native distribution coverage remains the package release gate. The candidate Worker helper executes the selected artifact's Blob or data URL bytes; only source-mode controlled imports compile worker.ts separately. The frozen published helper remains unchanged. yarn test:danmuku-types checks the unchanged npm root, the optional accurate /runtime entry, implementation assignability and semantic editor declarations. yarn test:danmuku-types-package packs and installs the packages outside the workspace, verifies exact member bytes and frozen offline reinstallation, and checks historical and current compiler consumers. These type checks do not replace the native browser suite or final distribution acceptance.

Documentation pipeline regressions

node --test test/documentation-pipeline.test.js covers the frozen old translator's delete-before-request, broken fence repair and exhausted 429 behavior; the new draft workflow checks failure, worker cancellation/join, reviewed apply, stale inputs, path escape, rollback and concurrent edits. Existing source Markdown round-trips and the offline source corpus are checked against actual inputs. A loopback HTTP server verifies a stalled response body times out. Mock responses test bounded retries and invalid data; no paid translation is performed and these tests do not certify English prose quality. Included in test:node.

Site build regressions

yarn test:site-build covers i18n compile failure, staged replacement and rollback, concurrent build/output conflicts, retained backups, actual language dictionaries and globals in four module modes, and a real Yarn documentation fixture child. Use this Yarn script instead of yarn node, which does not supply the lifecycle environment used to locate the pinned Yarn executable. The Markdown test uses the installed VitePress renderer to reproduce random code-group IDs and verify stable IDs, uniqueness and label associations. Built-page browser checks live in test/browser/document-site.spec.js; their editor destination is intercepted.

Desktop editor regressions

yarn test:site-editor covers frozen old TS execution/file error behavior, latest-run ordering, library/compiler failures, FileReader settlement and storage denial. test/browser/site-editor.spec.js uses the actual Monaco 0.30.1 worker and original vendor assets for typed Run, Ctrl-S, imports, preferences, declaration failure, language readiness and model disposal. Console Error objects are inspected through their arguments: Firefox may serialize their display text as only Error.

Multiple subtitles core combinations

test/browser/multiple-subtitles-combinations.spec.js uses verified old core and plugin bytes plus candidate source builds. It covers native caption timing, VTT/SRT selection, offsets/fullscreen/source changes and multi-instance URL ownership. Old1.0.0/1.1.0/1.2.0 factories run unchanged on the candidate core. Explicit single-active-cue probes record historical and candidate overlap defects; their passing assertions do not mean simultaneous captions are compatible. Candidate paused Firefox offsets are corrected by CORE-SUBTITLE-OFFSET-01. Old Firefox hosts retain explicit defect observations; their behavior is not silently counted as correct. Task09 classified the apparent WebKit source-switch caption loss as an old-host seek landing near zero, reproducible without any plugin. See refactor/changes/2026-09-14-PKG-MULTI-SUB-05-combinations.md before changing expectations. Reports include native active cues and rendered text. The new frozen 5.1.2 core route verifies archive integrity and member hashes before serving.

subtitle-offset.spec.js checks current/actual published core offset behavior, including synchronous native/public/rendered captions, identical cue objects, unchanged media time, disabled mode and resumed playback. The published Firefox branch asserts the original incorrect captions; the candidate must have the correct membership and order. subtitle-offset-native.spec.js loads a real HTML track without creating an Artplayer instance to compare native timing edits, cue reinsertion and mode toggling. It records actual time within the required interval, not a fabricated millisecond-exact seek landing. These diagnostic observations do not claim every native engine strategy is correct.

multiple-subtitles-switch.spec.js records source Promise settlement, native currentTime writes/events and actual seek landing with no plugin, candidate and published1.2.0. Old first-seek misses are retained as observations; every candidate first seek must land. The caption combination fixture separately waits for native source seeking to finish and for fullscreen-enter notification before testing subsequent commands. It retains exact caption and enter/exit order assertions; do not add delays or plugin retries to cover old core readiness behavior.