fix(#512): latch a late init ACK instead of losing it between retry windows

[debugmcpdev]

Fixes #512.

What

sendInitWithRetry registered its init-received listener per attempt and removed it when that attempt's window expired. An ACK landing during the inter-attempt backoff sleep fired into a void and was lost forever. Attempt 1's window is 500ms, so any worker boot slower than that fell into the gap — and a dry-run worker (which acks, finishes, exits by design) then converted the miss into Proxy process not available on every remaining retry, failing the launch ~16s later even though the worker did everything right.

Root-caused live while validating #501: with a fresh js attach to the server, worker spawn→ACK latency measured 669–842ms vs 338–427ms unattached (0/6 vs 6/6 cycle success — the boundary is exactly the 500ms window). The ACK provably reached the parent's ChildProcess in every failed cycle, captured by an evaluate_expression-injected monkeypatch on ChildProcess.prototype.emit after strace's ptrace overhead made the race vanish. Nothing is inspection-specific — a loaded CI host can add the same ~300ms, so this likely explains a family of flaky Failed to initialize proxy after 6 attempts failures.

Changes

• One latch listener spans the whole retry sequence; both the attempt window and the backoff sleep end early the moment the ACK arrives (checked again at each loop top).
• A failed send skips the pointless ACK window — nothing was delivered — preserving the previous retry cadence for the send-failure path.
• Once the worker has exited without acking (lastExitDetails recorded), the loop fails fast with the existing detailed exit message instead of burning the remaining ~15s of retries: after exit, no ACK can arrive.

Testing

• proxy-manager.handshake.test.ts: the test that pinned the drop ("retries when acknowledgement arrives after the first timeout") now asserts the latch (resolves on a single send); new tests for later-attempt-only ACK and the fast-fail path (2 sends, detailed message, no 15s burn).
• proxy-manager.start.test.ts: the exhaustion test now produces its exit details realistically (worker exits mid-retry) and asserts fast failure with the details attached.
• Suites run ~6× faster as a side effect of the fast-fail (17s → 2.9s for the two files).
• Live: integrated with js attach to a Node process that forks children strands those children (second startDebugging target ignored, child waits forever) #501, the self-attach stress harness went 0/6 → 6/6, with failing-window launches completing in ~1s (ACK latched during backoff).

🤖 Generated with Claude Code

https://claude.ai/code/session_01NgA5kHzkhFZLqbWNZRTsVG

[codecov]

Codecov Report

❌ Patch coverage is 93.02326% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/proxy/proxy-manager.ts	93.02%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!