feat(cli): nativeproof inspect — selector discovery for the current screen

[BenSheridanEdwards]

Summary

• Implements step 2 of DX: failure triage + selector discovery (kill the read-the-XML-and-guess loop) #23: nativeproof inspect (with --ios/--android/--project selection) starts a session with the configured project (noReset so app state survives), reads the current page source, and prints the candidate native.* locators a spec author would write — semantic roles with accessible names first, then every visible text (the exact alternation getByText matches), then test ids. The authoring-time counterpart of the did-you-mean failure hint: nobody reads XML and guesses strings.
• The extraction is a pure, unit-tested function (selectorSuggestions in src/inspect.ts); values are decoded, deduplicated, capped at 60 chars, and iOS names that merely mirror the label are dropped. Session capabilities are shared with the runner via a new projectCapabilities() (pure refactor of buildWdioConfig's spread).

Proof

• npm run check
  • Passed on feat/inspect.
• npm test
  • Passed: 158/158 (new: Android and iOS suggestion extraction — including the placeholderValue trap and label-mirroring names — dedupe/length rules, parseArgs/help wiring).
• Generated project/device proof, or N/A:

  • Device proof ran: nativeproof inspect --android against a live Android 15 emulator on the Settings home screen printed 41 candidate locators; the verbatim output is committed at docs/proof/inspect/android-settings-inspect.log. Sample:

    nativeproof inspect — 41 candidate locators on the current android screen
      native.getByRole("image", { name: "Profile picture, double tap to open Google Account" })
      native.getByText("Network & internet")
      native.getByTestId("com.android.settings:id/homepage_title")
    

  • Screenshots: Not applicable; the deliverable is terminal text, committed verbatim as the log above.

Risk

• Low. Additive command; the only shared-code change is extracting projectCapabilities() from buildWdioConfig (same spread, covered by existing config tests).

Notes

• Full device coverage: Android ran live above; iOS extraction is unit-proven on real XCUITest source shapes (the session path is platform-agnostic wdio).
• v1 launches the configured app and inspects its first screen; attaching to an arbitrary mid-flow screen stays with the (stretch) trace work in DX: failure triage + selector discovery (kill the read-the-XML-and-guess loop) #23.

PR Proof Law

• I followed .agents/DEFINITION_OF_DONE.md and .agents/skills/pr-inline-screenshot-proof/SKILL.md.
• Screenshots are committed and embedded inline with ![alt](...png?raw=1), or the proof section says Not applicable with the technical reason.
• The PR body has no bare screenshot links, local paths, relative paths, or proof placeholders.