js repl: surface thrown errors, skip action read-back on macOS

[hyprcat]

Follow-up to the numbers in #78: keeps the adapter-over-native-MCP design, fixes the two things that made it slow and fragile.

1. Thrown errors settle the call

A throw inside js code (sync, or after an await) never reached the REPL eval callback: Node prints it and moves on, so the call sat until timeout_ms (30 s) and then reset the session, taking app bindings with it. The session now settles the active evaluation from the REPL's error path (handleError on Node 26, the domain error event on Node 22) and returns isError with the message. Bindings survive; only timeouts reset.

Also: errors thrown inside the REPL context are from another realm and failed instanceof Error, which produced Error: Error: ....

2. Actions skip the read-back when a program drives them

Every native action settled ~150 ms and returned a fresh AX snapshot that app.click() etc. discarded. New OPEN_COMPUTER_USE_ACTION_READ_BACK=0: macOS actions return a short ok, no settle, no snapshot. All action-result sites route through one actionResult(for:); the settles that exist only for the read-back go through pauseBeforeReadBack(_:). Read per call so the app agent's per-request env forwarding applies. The adapter sets the variable when launching the native runtime. Default nine-tool behaviour is unchanged.

Calculator.app, one js call per row, 5 reps, median ms, display verified:

scenario	before	after
clear×2 + click "7" + read	2295	402
clear×2 + click "7" ×10 + read	6469	405
thrown exception → error returned	30000 + session reset	0, bindings intact

Raw native click with the flag: 270 ms → 4 ms. What's left in the rows above is the adapter's own getAXState() reads.

Linux and Windows runtimes ignore the variable and still read back after each action; the same switch would go in their actionResult paths and runtime scripts. Left out so you can decide whether you want it there.

Validation

• node --test scripts/node-repl/open-computer-use-repl.test.mjs — 13/13 on Node 26.4 and Node 22.23 (new test fails with a 1 s timeout before the fix)
• swift test — 168 passed, 1 opt-in live test skipped
• ./scripts/run-tool-smoke-tests.sh — native nine-tool and cursor-idle smoke passed (default read-back)
• make check-docs, git diff --check
• Bench script: https://gist.github.com/hyprcat/8997d42d2f226fbac9ce13e49b46ec27

[iFurySt]

@hyprcat Thanks for the fix! One edge case may still be problematic: if a completed js call schedules an async callback that throws later, the shared rejectActive can reject the next active call instead, and subsequent calls may hang until timeout. I reproduced this on Node 22 and 25 with setTimeout(() => { throw new Error("late") }, 50). Could you confirm and add a regression test for this case before we merge?

[hyprcat]

@iFurySt confirmed on Node 22 and 26. A's late throw rejected B. The same bug had a second symptom: B kept running after the early rejection and its nodeRepl.write landed in C's output.

Fixed in 32b8ec0. Each evaluation now runs inside an AsyncLocalStorage owner, and timers and promises carry it. The error path settles only the call that threw, and nodeRepl output is refused to code whose call has already returned. A late error from a finished call is dropped, same as the REPL did before this PR. If you'd rather surface it in the next call's output as a note, that's a small change.

Regression test (<Session>: a late throw from a finished call does not reach the next call) runs on both PersistentJavaScriptSession and WorkerJavaScriptSession. It fails on the previous commit and passes on Node 26.4 and 22.23, with all 21 scripts/node-repl tests green on both. I didn't hit the follow-on hang in my runs, but the misattribution it would come from is gone.

Also rebased onto main after #80. The only conflict was the release-notes row, and I kept both.

[iFurySt]

cool. lgtm. thanks!