feat(studio): AI chat context is opt-in (last run, widget tree), redacted; backend for a selector fix

[pbertsch]

Summary

• The Studio AI chat sends the open file with every message, as before. The failures of the last run and the device's widget tree are sent only when the workspace turns them on. Both switches are off by default. They live in probe.yaml under studio.ai_context (run_results, widget_tree) and in the Workspace settings overlay, under AI chat context.
• The Go side reads the switches from the workspace's probe.yaml for every chat turn, and it fails closed. With no workspace, an unreadable file, or a switch that is off, nothing beyond the open file is sent, and the device is not asked for its tree.
• While either switch is on, a line above the chat input says what is sent: "Sending the last run and the screen to the AI provider", or the one-part variant.
• Redaction happens before anything is sent. An editable field that is obscured, or whose key names a secret (password, PIN, token, OTP, CVV, API key, credential), loses its key and is marked. Credential-looking text (sk-…, ghp_…, Bearer values, long hex and base64 runs) becomes [redacted] in the tree and in the run text. Each part is cut to its size limit and marked when it is cut.
• SuggestSelectorFix (backend only) asks for one ProbeScript line that replaces the last failed step. It needs the run-results switch and refuses without it. Nothing is applied. The pane button, and the regenerated Wails bindings for it, are a follow-up.
• The Workspace settings overlay now reads and writes the workspace root's probe.yaml, not the folder currently browsed in the file pane. The toggles and the chat line therefore use the same file as the Go side. The fields the overlay already managed are unchanged.

What the widget tree contains today (please review)

The agent's probe.dump_tree returns widget types and ValueKeys. It does not return the text typed into fields, so typed text is not in the tree today. The agent writes no "obscured" marker either, so the obscured rule has nothing to match yet; the secret-name and credential rules do apply. Writing an obscured marker into the agent's dump is the follow-up that would make the obscured rule effective.

Test plan

• go test ./studio/ on the whole package, merged with origin/main, with the real frontend build embedded. The new tests cover: off by default (no probe.yaml, no key, no workspace); neither part is sent and the device is not asked when both are off; each switch sends only its own part; both on sends both, redacted; no secret reaches the prompt; SuggestSelectorFix needs the run switch; redaction of common credential shapes; redaction of obscured and secret-named editable fields; probe.yaml save and load keep the switches and the rest of the file, and switching both off removes the section; a studio key that is not a map is refused and the file is left alone
• go vet ./studio/, and gofmt for the new Go files
• npm run build (tsc and vite build) on the merged tree
• Not run against the live Anthropic API. The tests use a fake provider and make no network calls
• Not run in the Studio window: the settings overlay and the chat line were not exercised by hand

🤖 Generated with Claude Code