fix: preserve long-running cleanup requests

[lindixu6-hash]

Summary

• stop treating a fully consumed Node request body close event as a client disconnect
• abort the fetch-style request only on a genuine request abort, premature response close, or socket failure/close
• derive the Node HTTP idle timeout from the longest profile-cleanup deadline with a 30-second transport grace period
• cover normal body completion, genuine disconnect deduplication, completed responses, and timeout ordering

Fixes #285
Fixes #286

Validation

• HOME=/tmp/opencode-mem-lifecycle-test bun test — 453 passed, 1 skipped, 0 failed
• bun run typecheck — passed
• bun run build — passed
• targeted Prettier check — passed
• git diff --check origin/main...HEAD — passed
• Node.js v26.8.1 lifecycle probe reproduced the report: a fully consumed POST emits req.close with req.complete=true while res.writableEnded=false; the new guard ignores that event
• GitHub package-smoke matrix — Ubuntu, Windows, macOS Intel, and Apple Silicon all passed

Boundary

The regression tests exercise the extracted lifecycle policy deterministically. The Node.js 26 probe verifies the runtime event ordering that triggered the original empty response.

[amandeavor]

LGTM! Tested locally on Windows (Node v24.16.0 / Bun 1.4.0):

• \�un test tests/web-server-request-lifecycle.test.ts tests/ai-cleanup.test.ts\ (12/12 passed)
• \�un run typecheck\ (\ sc --noEmit\ passed cleanly)

The extraction of \�ttachNodeDisconnectHandlers\ cleanly handles the Node 26
eq.close\ event order after full body consumption, and coupling the transport timeout (\NODE_HTTP_IDLE_TIMEOUT_MS) to the longest cleanup deadline plus grace period prevents premature aborts on large profiles.