Rerun the Trailhead benchmark from agent-created worktrees

[janicduplessis]

Problem

The first benchmark started both agents inside pre-created worktrees. The Stim arm also loaded Expo skills. The run did not test a warm main checkout followed by an agent-created isolated worktree.

Acceptance criteria

• Build the main checkout before either timed run.
• Start each fresh agent in the main checkout.
• Require each agent to create its own isolated worktree.
• Allow only Stim and Agent Device skills in the Stim arm.
• Disable all skills in the control arm, including Expo skills.
• Record commands, elapsed time, cache behavior, retries, and screenshot or recording evidence.

[janicduplessis]

Benchmark run 20260829T2244Z is complete.

• Control time to proof: 672.471 seconds.
• Stim + Agent Device time to proof: 206.468 seconds.
• Stim was 3.26 times faster to proof.
• Control retries: 10.
• Stim retries: 0.
• Both arms hit the same 19cb61 native artifact.
• Both arms passed lint, type-check, UI verification, proof capture, and cleanup.

The control prompt had no skills, but the agent discovered the installed stim and agent-device binaries. The result measures skill and workflow value, not binary availability.

The run found two follow-up issues:

• Wait for iOS Simulator boot completion before launch #109 waits for full CoreSimulator boot completion before launch.
• Make the fast worktree and Agent Device path explicit in the skill #110 makes --carry-ignored and the Agent Device Metro binding explicit.

One initial Stim attempt was invalid. Its isolated home contained the wrong prepared fingerprint. The resulting cold build exhausted disk space. The corrected run used the same artifact as control.