feat(b-roll): add video-add-b-roll skill

[ZhangchengCharles]

What changed

Adds video-add-b-roll, a new skill that places selective, transcript-timed visual cutaways into an understood Project Protocol V1 sequence.

• New skill — SKILL.md playbook, reference/broll-rules.md, an example plan, the local candidate review page, and five scripts: broll_plan.py, pexels.py, build_review_page.py, normalize_broll.py, check_broll.py, with a 154-test suite.
• Acquisition — Pexels search/download plus import-local for your own or externally generated footage. Every candidate is frozen with a real probe, byte count, SHA-256, and provenance; downloads validate redirects, hosts, size, and media before atomic publication, and .part recovery is owned by the downloader.
• Two-stage review — an exact-candidate review that must decide every shot once, then a completed visual review bound by SHA-256 to the plan, stills, contact sheet, boundary reel, machine summary, and final delivery. Human and explicitly delegated Agent modes are both valid; neither can be fabricated from silence.
• Shared protocol — projectlib.py gained b-roll plan and receipt validation, dependency-revision coverage, and image-sequence overlay checks. The operation adds a b-roll track and changes video pixels only, leaving timeline, geometry, and audio untouched.
• Docs — registered the skill in CLAUDE.md/AGENTS.md and the README, and documented the repo's runnable test suites and check_* harnesses, which were previously undocumented.
• Also on this branch (predates the b-roll work): the video-rough-cut → video-cut rename and the ClawHub publish workflow.

Why it changed

Talking-head, interview, and explanatory videos need a small number of evidence-backed cutaways. Doing that by hand means unverifiable provenance and no record of who approved which shot. This makes the shot list a reviewable, diffable plan and refuses to guess: missing, weak, or downloader-rejected media becomes skipped, and an approved all-skipped plan is a valid no-op rather than an excuse to drop in generic filler.

How it was checked

• python -m unittest skills/video-add-b-roll/tests/test_broll.py — 154 tests, OK
• skills/video-cut/tests/test_inspect_bounds.py — 6 tests, OK
• skills/video-edit-compare/tests/test_make_compare.py — 1 test, OK
• check_protocol_extensions.py plus the cut and captions protocol checkers — all passed
• Confirmed the supported Pexels endpoint is api.pexels.com/v1/videos/search (the unversioned path is deprecated) and that both routes returned identical live payloads before switching
• Fixed a doc/code break found while reviewing: complete_visual_review() required a fifth comparison argument that the playbook no longer passes, which raised TypeError for anyone following SKILL.md. Since SKILL.md is the spec, the code, compiler contract, and fixtures moved to match.

Notes for review

Two known issues, both pre-existing and deliberately left out of scope:

1. skills/video-to-shorts/scripts/check_review_ui.py has 2 failures (fake_render() missing video_preset). They reproduce identically with this branch's changes stashed.
2. video-add-b-roll and video-add-content-cards both publish under review/03-. The README now reflects that real behaviour rather than hiding it. Renumbering isn't a docs change — b-roll has 54 references to review/03-b-roll, and those paths are SHA-256-bound in review receipts, so it needs its own commit.

The branch is named docs/ but the diff is mostly feat; the branch predates the feature work landing on it.