test(e2e): prove AskUserQuestion against the real CLI, not a reading of the SDK

[ElbertePlinio]

Closes the gap I flagged when #364 merged: it shipped with thorough unit coverage at every layer, and no way to know it worked.

Nothing in this repo reaches the CLI:

tier	what it mocks
claudeBridge.test.ts	mocks the SDK, puts a zero-byte claude on PATH
Rust bridge tests	spawn #!/bin/sh fakes
VRT	runs against tauriMock.ts
bun run e2e	mocks invoke

So the one contract that actually matters — that returning {behavior: "allow", updatedInput: {...input, answers}} makes the tool report answers rather than "The user did not answer the questions." — was asserted against my reading of the SDK, never against the SDK.

What this does

Drives the real claude binary through the real bridge, forces an AskUserQuestion, answers it through the same approve op the GUI uses, and asserts the tool's own result text.

Verified load-bearing the only way that counts

answers branch enabled  → PASS: AskUserQuestion was answered and the tool accepted it
answers branch disabled → FAIL: the tool reported that the user did not answer

Disabling the branch in permissionResultForDecision reproduces the real refusal string from the live tool; restoring it passes. That is the first direct evidence #364's fix works.

Gating, per the device tiers beside it

• Unset gate → SKIP, exit 0. CI and bun run e2e stay green without credentials.
• Gate set but claude missing → FAIL FAST. A silent pass here would be worse than no test.
• Costs a real API call; pinned to a cheap model (PICKFORGE_E2E_CLAUDE_MODEL, default haiku) per the repo's dogfooding convention in CLAUDE.md.
• bun run e2e:claude added; tests/e2e/README.md documents the tier.

One finding worth the README entry

Wait for a raw result message, not turnClosed. The bridge emits turnClosed only when the whole query iterator ends — a session stays open for follow-up sends — so waiting on it hangs until the timeout even though the turn succeeded. That cost a debugging cycle and would cost the next person one too.

Incidental confirmation for #365

The live stream shows system:status and system:thinking_tokens frames arriving and being dropped by the parser's dispatch — direct evidence for the premise behind #392, which handles system/status.

Validation

• PICKFORGE_E2E_CLAUDE=1 bun run e2e:claude — PASS against the live CLI
• bun run e2e — PASS (device tier skips, as designed)
• bunx tsc --noEmit clean; bun run lint clean

Refs #364