docs(skills): add pre-build gate to changelog-video SKILL.md

[jrusso1020]

What

Adds an explicit Step 0 above the existing pipeline in .claude/skills/changelog-video/SKILL.md (and the mirrored .agents/skills/changelog-video/SKILL.md) that spells out the concrete pre-build file copies before any composition HTML is written:

• cp <SKILL_DIR>/assets/fonts/*.woff2 project/assets/fonts/
• cp <SKILL_DIR>/assets/bgm.mp3 project/bgm.mp3
• ffmpeg … re-encode of <SKILL_DIR>/assets/bg-pattern.mp4 to the total duration
• cp <SKILL_DIR>/examples/master-skeleton.html project/index.html
• Read references/build-spec.md end-to-end BEFORE writing content

Also extends the anti-patterns table with the three failure modes hit on this week's Jul 13-20 weekly run:

• Starting from a prior video's index.html
• Hand-crafting @font-face / WebGL shader / custom BGM
• Delivering without a CloudFront invalidation after the S3 replace (distribution E2BSLVSZ7FG3U0)

Why

The previous SKILL.md said "follow references/build-spec.md exactly" in step 5. That lets an agent read the build-spec after already writing composition HTML on a wrong scaffold — which is exactly what happened on the Jul 13-20 run: three off-brand video iterations (WebGL shader background, JetBrains Mono fonts, blue accents) before finally starting from master-skeleton.html. Jake (skill author) called v3 "trash" — the assets, fonts, and scaffold ARE the brand, not the SKILL.md prompt on its own.

The three added anti-patterns are the concrete traps: agents keep starting from similar-looking prior-project templates unless the skill explicitly bans it and points at the exact cp commands.

Test plan

• bun scripts/lint-skills.ts — 31 skill file(s), 0 issues (fixed inline-backtick > false positives on the first attempt via rewording)
• node scripts/check-skill-mirror.mjs — 24 files byte-identical across .claude/skills/ and .agents/skills/
• bun run lint — 0 warnings, 0 errors
• bun run format:check — all matched files use the correct format
• bunx fallow audit --base origin/main --fail-on-issues — ✓ No issues in 2 changed files

Content-only change; no scripts or assets touched.

Follow-up

Home's weekly cron (cron_dfd99766) — I've sent the corresponding cron-prompt update (via send_to_teammate) so next Monday's autonomous run picks up the pre-build gate before this PR merges. This PR is the durable skill-level fix; the cron update is the specific-cron-flow fix. They're additive.

[vanceingalls]

LGTM.

What I checked:

• Both mirrored SKILL.md diffs are byte-identical (per PR body + independent diff read). Consistent with the mirror gate.
• All referenced paths exist on main: .claude/skills/changelog-video/assets/{fonts/,bgm.mp3,bg-pattern.mp4}, examples/master-skeleton.html, references/build-spec.md. Step 0's cp targets are real.
• Ran the anti-patterns against the failure narrative: (1) prior-video index.html reuse → covered by the master-skeleton cp; (2) hand-crafted @font-face / WebGL / BGM → covered by assets/ verbatim copy; (3) missing CloudFront invalidation → orthogonal delivery-side trap, worth calling out separately.
• ffmpeg command's bg-pattern-<TOTAL>s.mp4 output name matches the assets/bg-pattern-<dur>s.mp4 convention in the existing "Project layout" section.
• No conflict with existing step 5 ("Build") — Step 0 supplies the scaffold + Step 5 fills into it; the framing paragraph makes that division explicit ("step 5 fills placeholders inside the already-copied project/index.html").

One nit (non-blocking).

The Step-0 explanatory paragraph enumerates placeholders as <RANGE>, <TOTAL>, <CUT_N>, <DUR_N>, scene bodies. <RANGE> (3×) and <TOTAL> (7×) exist literally in master-skeleton.html, but the actual per-cut tokens are <CUT1>, <DUR1> (indexed) — and the skeleton also uses <THEME> + <HEADLINE> which the paragraph doesn't mention. Reads fine as metasyntactic (_N = "for the Nth cut") but a future agent doing a literal grep would come up empty on <CUT_N>. If you want to tighten this in a follow-up: swap for <CUT1..N> / <DUR1..N> and add <THEME>/<HEADLINE> to the "scene bodies" phrase. Not worth blocking on.

Ship it — the durable fix that would have prevented three off-brand iterations on the Jul 13-20 run.

Review by Via