fix: death-aware, retryable js-debug child adoption; remove dead configDone deferral

[debugmcpdev]

Fixes #248
Fixes #249
Fixes #250

First of two PRs for the cluster:js-child-session backlog chunk (#247 follows separately).

What changed

#250 — remove dead deferParentConfigDone machinery (commit 1). The 1500ms parent-configurationDone deferral could never arm on any real path: its only writer runs after createChildSession resolves, but the js-debug handshake sends parent configurationDone (step 4) before launch/attach (step 5), so no child can exist when the parent configDone goes out. Even force-armed, nothing ever resolved the deferral early — it always burned the full timer. Deleted along with the equally dead AdapterPolicy.shouldDeferParentConfigDone() and the DapClientBehavior.deferParentConfigDone flag from all eight policies (−233 lines).

#248 — death-aware adoption (commit 2). A child target dying mid-adoption used to burn every remaining step's full timeout (worst case ~400s in the attach retry loop; ensureChildStopped alone could add 80s) and left the half-adopted socket orphaned at worker shutdown:

• New settle-once ChildDeathLatch (modeled on fix: js-debug launch barrier settles on debuggee death (fixes #242) #246's JsDebugLaunchBarrier) subscribes close/error right after connect; every adoption step is raced against it and aborts the moment the child dies.
• attachChild gains a 60s total deadline on top of per-request timeouts, with death checks before/after each retry.
• waitForEvent settles immediately on client death instead of waiting out its timer.
• The failure path now rolls back childSessions/activeChild — previously hasActiveChildren() latched true forever, silently rejecting every subsequent adoption — and shuts down the failed child client.
• MinimalDapClient.shutdown() now also invokes ChildSessionManager.shutdown() (previously unreachable from src), covering mid-adoption sockets never promoted via childCreated.

#249 — retryable adoption (commit 3). The policy-level adoptedTargets dedupe kept the pending-target id when adoption failed, so js-debug's re-sent startDebugging was answered {handled: true} without creating a child. The dispatch site (which already observes the rejection) now rolls the id back, and shutdown() clears the set.

Testing

• 7 new unit tests: close-mid-attach aborts promptly, death at step boundary, attach total deadline, failure rollback + successful retry, manager shutdown reachability, re-sent startDebugging retries, adoptedTargets cleared on shutdown. The death tests run in ~35ms where the old behavior hung 15s+ (test-timeout kills).
• Full gate: 2593 unit tests, 18 e2e smoke files, break-on-exceptions e2e 8/8, lint clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 95.50562% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/proxy/child-session-manager.ts	96.34%	3 Missing ⚠️
src/proxy/minimal-dap.ts	85.71%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!