LLM taps by ref instead of x/y, OCR-checked iOS labels, failed-trail report buttons

[handstandsam]

Summary

• The LLM taps by ref, not x/y. tapOnPoint leaves the default core_interaction toolset; offered beside tap, models picked coordinates and missed. Recorded tapOnPoint steps still decode and replay, and a target opts back in with tools: [tapOnPoint].
• iOS visible strings are checked against the screenshot. On macOS, report export runs Vision OCR over each iOS AXe capture, so an icon-only button's label is filed as accessibility text rather than drawn text.
• Report: failed trails get trail and replay buttons, and a configurable "Fix this category" launcher action on analysis pages.
• Subprocess tools keep their full JSON Schema, so nested arguments (e.g. overrides[].value) coerce to the declared type on dispatch.
• Event diffs summarize list-shaped events by their items instead of [object Object].
• Host-only tool check accepts a requires-host-tools.txt allowlist from repos that vendor the framework.
• Devlogs: Trailblaze vs Maestro (TypeScript tools, subflows, Maestro MCP), and why the UI moved from Compose to a TypeScript web app; older Compose entries marked superseded.

🤖 Generated with Claude Code