feat(agentic): migrate recipe-performance and add paired device measurement - #225

Migrate the existing recipe-performance workflow into the agentic domain and add Mobile bottleneck diagnosis and paired proof for devices, simulators and emulators.

• Preserve the Extension collector, evidence rules and checks.
• Teach thread classification, CPU sampling/flame graphs, bounded CDP transfers, whole-subtree elimination and alternating production-sampling proof.
• Add engineer-facing prerequisites, one-line usage, expected outputs and a capture-time budget. Remove the separate Mobile measurement skill.

Validation: seven focused kit checks, Extension collector checks and repository CI pass. Retained raw reports reproduce byte for byte. A failing-first deep-stack regression and real-profile replay cover CPU rendering without interpreter tuning.

Blind iOS-simulator validation on this head produced a lazy funding-hook optimization. Four alternating restarted blocks, n=10 per state per direction: first-JS-frame delay improved 17.7% Long / 17.9% Short; opening improved 14.6% / 16.2%, with block separation and better p75/readiness. 246 focused product tests and the 39-step recorded control proof pass.

Long's post-open maximum JS gap worsened; total post-open blocking overlapped. No native/compositor FPS claim. The formal product visual gate remains pending; the local demonstration is unpublished. A self-contained HTML results page includes the actual flame image, code diff, wall time and comparison with the prior physical-device result.