Add selected research-to-delivery skills, artifact checks and a worked example

[jdavis-software]

Selected scope

Implement the six useful additions approved from the twenty-repository review as original requirements-first work. Leave optional graph/search engines, overnight controllers, paid user-study integrations, slide exporters and scroll-video systems out of this release.

Six new skills: Time-windowed Research; Technical Debt Triage; Evidence-backed Visual Explanation; Product Video Storyboarding; Media Transform Verification; Search Discoverability Audit.

Six focused refinements: Multi-source Research, Bounded Context Assembly, Context Checkpointing, Work Packet Planner, Agent Runtime Adapter Qualification and MCP Adapter Design. These add claim/excerpt states, scoped retrieval, evidence-bound lessons, reuse-before-rebuild, independently checked protocol boundaries and invocation-time authorization without duplicate catalog entries.

New original utility: Publicationcheck, a Node wrapper around dependency-free Python modules. It checks bounded local research/debt/storyboard/explanation records against selected file hashes and excerpts; audits a static build's canonical URLs, indexing directives, sitemap/internal links and JSON-LD syntax; and explicitly probes/decodes a small owned local media file with installed FFprobe/FFmpeg. It does not crawl, launch agents, render video, authenticate approval metadata or decide semantic truth.

Website: Research to Delivery bundle, Design & Communication category and an original worked example with source-linked synthetic stages, relationship text, a timed storyboard, explicit future-feature labeling and executable-check guidance. No remote assets or external rendering engine are needed by the example.

Inventory: 88 entries / 79 original skills / 9 tools / 13 bundles / 14 categories. Eighteen new synthetic host-evaluation inputs remain not-run; 219 accumulated domain scenario inputs are separate from executed tests. Existing bundle resolution includes the actual companion modules and documentation.

Final combined-candidate validation — passed

Final head 544ca9b, source tree bcf486dd529d990b3e86df1b023d1a8d6ac345c4, tested as combined candidate b2fc8905f3fcf063931995d05996e1f1f06232ea6 against main cb53b54.

Successful final validation:

Check	Actual result
Frozen dependency installation; catalog, categories, bundles, sources and README	Passed
Node tests	256 passed, zero failures/cancellations/skips
Sourcekit Python regressions	80 passed
Publicationcheck Python tests	70 passed, including real bounded child-process tests
Astro/TypeScript	Zero errors/warnings; two existing unused-symbol hints
Production build	Passed
Publicationcheck static-site audit	123 HTML pages inspected, no findings under its declared checks
Original record demonstration	Four valid cases pass; four corresponding deliberate defects detected
Actual FFmpeg/FFprobe smoke	Valid synthetic media decodes and passes; missing audio and wrong duration rejected; input hashes unchanged
Desktop/mobile Chromium browser and request cases	50 passed, zero unexpected/flaky/skipped
Existing Sourcekit live public reads	Two approved public HTTPS reads passed

The media smoke uses the distribution FFmpeg package on the disposable Ubuntu runner, recorded as 7:6.1.1-3ubuntu5, not an installed upstream skill package. It generates only short synthetic local clips. Codec execution, missing-audio and duration checks were actually exercised, not inferred from a process exit code.

The final publication artifact was downloaded and verified at SHA-256 20b880bfb1e576ccf4d1d5dbad1a42944c003dda3986720bda01dec35ec1fe46 before inspecting records/site/media JSON. Its four seeded record defects are an old republished event presented as current, a confirmed debt finding without reproduction evidence, an unsupported relationship, and a planned feature presented as available. These tests validate references and consistency, not the underlying truth of arbitrary supplied records.

The final browser/build archive was verified at SHA-256 d884ef6c4bbc5386e9f4ea0180e39e09cf46470a1e8fa72a428e418e3c84c642. Its report confirms 50 expected passes, no unexpected/flaky/skipped cases, and build-info identifies the combined candidate and inventory above.

Browser and visual review

Existing Playwright workflow ran desktop Chromium 1440×1000 and iPhone 13 mobile emulation; this is not physical-device/Safari or live-agent testing. The new journeys exercise bundle → worked example → relationship evidence → Publicationcheck, plus canonical skill sources, dark/light theme persistence and no-JavaScript content. Existing catalog search, filters, category counts and mobile navigation remain covered.

Actual final desktop and mobile screenshots were reviewed in both themes. The intentional additions preserve the existing restrained lime/dark/light identity, use three-column desktop and stacked mobile stages/storyboards, retain an accessible relationship-text disclosure and mark the fixture and planned feature explicitly. No page-horizontal overflow was found by the browser assertions.

Issues found and fixed during qualification

• The initial runner lacked FFmpeg/FFprobe: qualification failed rather than skipping. Explicitly provisioned the system dependency in disposable CI and reran the real media smoke.
• A subsequently passing run executed only the old 44 browser cases because Playwright uses an explicit testMatch registry. Inspection caught the omission. Added delivery.spec.ts and reran; the final 50 cases include all six new desktop/mobile checks.
• Added actual subprocess tests for stdout/stderr handling, nonzero exits, output limits, open-pipe timeout and a child closing both pipes before hanging. Kept invalid/unknown duration out of successful media evidence.

Limits, preservation and publication gate

No npm dependencies or dependency-lock edits, copied upstream skill bodies, style databases or assets, model/provider spending, new recurring automation, original-content license choice or private DriftGate/AvatarOps repository/config/MCP changes. Temporary source-transfer workflows and patch fragments are absent from the final tree.

Record checks establish bounded byte/reference/date/state consistency, not semantic entailment, source independence, actual defect reproduction, rights or authenticated approvals. Site checks establish the inspected local technical conditions, not live indexing/ranking or rich-result eligibility. Media checks require trusted owned inputs, bounded local protocols/process output/time, and do not constitute an OS sandbox or perceptual/caption-sync assessment. Skills retain experimental host-effectiveness labels.

Canonical contracts: docs/PUBLICATIONCHECK.md. Reproducible fixtures: examples/publicationcheck/. The merged-main workflow will rebuild/test and verify the exact public revision and inventory before publication is reported complete.