feat: add feature-demo skill

[klepfish]

New Skill

Skill name: skills/feature-demo

Summary

feature-demo records a Playwright-driven demo video of an in-app feature flow (signup, onboarding, a settings flow, a dashboard interaction) and delivers a polished MP4 — branded title card, on-screen subtitles, an animated mouse cursor, and optional ElevenLabs TTS narration.

It complements the existing video skill rather than overlapping it: video covers AI-generated / programmatic marketing video (HeyGen, Remotion, Veo, …), while feature-demo is for real screen-capture walkthroughs of an actual running web app. The description cross-references video to keep the scope boundary clear.

What ships:

• SKILL.md — end-to-end workflow: gather audience/feature context → bring the local app up → write the recorder → record → optional narration → deliver.
• scripts/recorder.template.mjs — copy-and-edit recorder skeleton (BRAND / TITLE_CARD / FEATURE STEPS blocks).
• scripts/narrate.mjs — generic narrator that muxes ElevenLabs voiceover onto the silent capture (caching, overlap-avoidance, title card).
• references/step-modes.md — decision matrix for aligning narration with on-screen actions (before / during / after).
• references/local-app-gotchas.md — cross-cutting HTTP-localhost recording issues (cookie flags, CSP, demo-state reset).
• evals/evals.json — 6 evals exercising the expected behaviors.

Stack-agnostic — works against any local app (Laravel, Rails, Django, Next.js, …) that renders in headless Chromium.

Repo bookkeeping

• New skill versioned at 1.0.0 (matching how sms / prospecting entered the 2.x line).
• Bumped the marketplace release 2.3.0 → 2.4.0 and added a VERSIONS.md changelog entry + table row, following the one-minor-per-new-skill pattern.
• Ran node .github/scripts/sync-skills.js, which regenerated marketplace.json (count 43 → 44), plugin.json (version → 2.4.0), and the README skills table — so those match what CI would produce on merge.
• Kept the file cross-agent: no allowed-tools, no agent-specific tool calls (the delivery step is tool-agnostic), and install paths reference .agents/skills/… per AGENTS.md.

Validation performed

• bash validate-skills.sh → 44/44 skills valid, 0 warnings, 0 issues.
• node --check on both .mjs scripts; evals.json parses as valid JSON.
• Scanned the skill for secrets / internal references — none.

Checklist

• name matches directory name exactly
• name follows naming rules (lowercase, hyphens, no --)
• description is 1-1024 chars with trigger phrases
• SKILL.md is under 500 lines (292)
• No sensitive data or credentials
• Tested locally with AI agent (contributed from another project; validated here via validate-skills.sh, node --check, and the evals)