feat: add simulator verification skills

[thymikee]

Summary

Add focused iOS Simulator and Android Emulator skills for agent-device, and link them from AI Agent Setup.

Keep the new source-distributed skills out of the npm tarball by limiting the package allowlist to the existing packaged skills.

Validation

• pnpm check:affected --run in a clean temporary worktree (all runnable checks passed).
• npm pack --dry-run --json --ignore-scripts confirmed only skills/agent-device and skills/dogfood are packaged.

The normal worktree contains unrelated local .claude/worktrees fixtures with intentionally invalid YAML, so its full formatter cannot be used as validation for this docs/packaging-only change.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-10 19:16 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	2.05 MB	2.05 MB	0 B
JS gzip	665.2 kB	665.2 kB	0 B
npm tarball	802.2 kB	802.2 kB	+7 B
npm unpacked	2.80 MB	2.80 MB	+35 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	23.4 ms	24.2 ms	+0.8 ms
CLI --help	54.9 ms	54.6 ms	-0.2 ms

Top changed chunks: no changes in the largest emitted chunks.

[thymikee]

Exact-head review at 607becee9: not ready. The source-distributed skill/package scope is coherent, documented command shapes match the shipped CLI/help contract, and no device evidence is required for this skill/docs/metadata-only change. P2 validation gap: both new files add command-planning guidance, but there is no required help-conformance case or structural skill-contract regression for either one; the existing help-conformance test is hard-wired only to skills/agent-device/SKILL.md. Add focused iOS/Android cases that prove the skill selects the requested platform, starts a known app with foreground snapshot composition, grounds actions in current refs/selectors, verifies, and closes, and demonstrate the gate fails when that key guidance is removed or corrupted. Completed checks are green; Android/iOS smokes are still pending. No ready label applied.

[thymikee]

Exact-head re-review at 1487f873: the updated iOS/Android skill wording accurately matches the CLI contract, all checks are green, and removing the unused OpenAI metadata is fine. The prior P2 remains: this delta adds no help-conformance or structural skill-contract coverage. The existing harness is hard-wired to skills/agent-device/SKILL.md, so either new simulator skill can lose/corrupt its platform selection, foreground-open/snapshot, current-ref/selector, settle/type, verify, or close guidance without a gate failing.

Add focused iOS and Android conformance cases and a planted-red proof for each required skill contract. No device evidence is needed for this docs/skill-only change; no ready label applied.