agents: add performance-engineer (render-pipeline perf expert)

[jbrecht]

Adds a seventh subagent on the one review axis the existing six don't cover — "is it fast / efficient."

Why

code-reviewer (correct), qa-engineer (proven), designer (watchable), release-reviewer (safe to ship) — none own wall-clock render time or encode efficiency. With umami #133 about to regenerate its whole tutorial set, batch render cost is becoming real.

Guardrails baked into the prompt (so it's not generic "use -preset faster")

• Profile before prescribing — find whether time goes to TTS / record / post, then target the dominant cost; report cold-vs-warm-cache and single-vs-batch separately.
• Respect the hard constraints — works on every ffmpeg build (no assuming hwaccel; optional + fallback only), A/V sync (flash calibration) is sacred, no output/behavior change, and the record phase is content-bound by narration-first pacing (don't chase it).
• Quantify — every finding carries an estimated saving × the regime it applies to; safe wins separated from output trade-offs that need a product call.
• Advisory — measures and recommends; routes correctness risks to code-reviewer, watchability trade-offs to designer, and accepted optimizations through product-manager.

README updated (count 6→7, axes, reactive table, division-of-work, and the ffmpeg-args hand-off seam with the code-reviewer).

Note: these files live under .claude/ (a dot-directory), which dorny/paths-filter (0picomatch, default dot: false) does not match with **/*.md — so this PR correctly ran the full gate. The docs-only skip works for docs/** and root *.md; covering .claude/** config changes is a follow-up.

🤖 Generated with Claude Code