fix(interact): cdp_interact value-injection for Controller inputs (#336)

[Lykhoyda]

Closes #336.

Problem

cdp_interact corrupted react-hook-form Controller-wrapped inputs two ways:

1. Press → event object. press testID=<radio> called onPress({nativeEvent:{}}); for a value-bearing control whose onPress routes to the Controller's onChange, the event object became the field value → Zod "received object".
2. Digit-string → number. setFieldValue value="15112345678" reached setValue as the number 15112345678 → z.string() fails "Expected string, received number".

Root cause (verified)

• Bug post-edit-health-check hook fires false positives outside RN project and when app is uninstalled #1: injected-helpers.ts press block passed a synthetic event unconditionally.
• Bug post-edit-health-check still fires when app is not installed/running on booted simulator #2: not a coercion in the bridge — JSON.stringify preserves types and the helper forwards value verbatim. The trap is the value schema union admitting number, so a digit-string is emitted/parsed as a number. The number/boolean passthrough is an intentional feature (gh-126 asserts it), so the fix must preserve it.

Fix

• Bug post-edit-health-check still fires when app is not installed/running on booted simulator #2 — setFieldValue type-matches: if the injected value is a number and the field currently holds a string (formReturn.getValues(name)), coerce back to string. Number/boolean fields (current value number/undefined) are untouched; a getValues throw → no coercion. Result carries coercedToString.
• Bug post-edit-health-check hook fires false positives outside RN project and when app is uninstalled #1 — press gains optional value: when provided, onPress(value) instead of the synthetic event (built imperatively in ES5 to match the injected-helpers' deliberate no-spread style). No value → byte-identical to before.
• HELPERS_VERSION 32 → 33 (re-inject); index.ts schema descriptions updated.

Testing

• 7 new unit tests running the actual dist/injected-helpers.js in a VM against realistic RHF-shaped fiber trees (setFieldValue coerce/preserve/throw-safe/string-unchanged; press with-value/without-value/no-handler). Full suite 2537/2537.
• Reviews: multi-LLM plan review (Codex + Gemini, no blocking findings — applied the one advisory: ES5 press return) + a final whole-branch review.
• Device verification: deliberately skipped — the fix is pure JS value-handling logic with no external-runtime dependency (unlike bug: keyboard-occluded bottom-pinned taps land on the soft keyboard → flaky replays (iOS + Android) #356's maestro-runner case), the unit tests exercise the real helper JS, and the test-app has no FormProvider/press-value-radio target. Covered by unit + review + CI per maintainer decision.

Docs

Spec + TDD plan under docs/superpowers/{specs,plans}/2026-06-30-336-*.

🤖 Generated with Claude Code