[XXPermanentXX]

When telemetry stalls, flushTelemetry() currently stops waiting after its budget but leaves DNS/HTTP work alive, so a command can print its complete result and still take tens of seconds to exit. This change cancels that work within the telemetry budget and releases unused response bodies.

Related to #294. This PR is stacked on #293 (output-contract-review); its diff contains only this telemetry fix. After #293 lands, retarget or rebase onto main as needed. Keep #294 open until the published Windows binary has been verified.

Changes

• Remove the independent DNS preflight. Give each send a private cancellation signal and a 1-second deadline; flush cancels pending sends when its own budget expires.
• Release unused responses and finish the private send scope on completion, including Bun's native-fetch cleanup. Node uses an owned dispatcher with cancellable direct/proxy TLS connections; business/global dispatchers remain intact.
• Preserve event fields/encoding, batching, opt-out switches, and existing CLI proxy environment selection/fallback. Do not change business commands, business request timeouts, or runtime exit handlers.
• Timeout-cancelled events are not retried. Other transport failures can be retried from the in-memory queue on a later send in the same process; there is no cross-process replay.
• Add 11 unit tests and 15 real-process regressions; retain all existing tests. Extend the telemetry maintenance checklist with cancellation and actual-exit requirements.

Validation

• Final-tree related regression: 98/98 passed across 10 files, including telemetry, proxy, Ctrl-C, EPIPE, concurrent invocations, and output/error behavior.

• Final full offline run (Node 26.13.0, Bun 1.3.14, 2 workers): 8,280 passed / 297 gated / 2 failed, 401 files, approximately 692 seconds. This is not an all-green full run. Failures were the existing compiled-binary version check (macOS SIGKILL) and a Sandbox polling-timeout assertion receiving the lower-level request-timeout message.

• Isolated rerun of both existing suites: 59/61 passed; all 13 Sandbox tests passed, while two compiled-binary startup cases received SIGKILL. On unchanged telemetry baseline 6f0cd3d0, the same two binary SIGKILL failures reproduced; the selected Sandbox recovery cases passed. These paths do not invoke the changed telemetry send (direct HTTP probe / telemetry-disabled runner), and no existing test was altered to hide the failures.

• Normal pre-commit hooks passed: regenerated skill assets without adding unrelated changes, then ran vp check --fix. git diff --check passed; the committed contents exactly match the tested tree.

• Separate Node 26.13.0 / Bun 1.3.14 loopback probes: successful HTTPS CONNECT with proxy authentication, NO_PROXY bypass, and stalled CONNECT cancellation all passed (6 cases). Successful delivery retained the exact POST payload; the shared dispatcher stayed intact. Stalled CONNECT sends cancelled at about 1 second and both child processes exited naturally.

• Same-machine Bun 1.3.14, identical fake business gateway and 6-second stalled telemetry, three runs each: median process exit 6.322 -> 1.312 seconds for stalled headers and 6.330 -> 0.337 seconds for stalled body. Business JSON and exit code were unchanged. A separate 50-second stalled-header reproduction exited in 1.499 seconds after the fix, compared with 50.414 seconds before.

Follow-up for Windows verification

• Reproduce the original command on Windows 11 / PowerShell with the release binary and npm/Node entry. Compare telemetry enabled, DO_NOT_TRACK=1, and NO_UPDATE_NOTIFIER=1 independently, recording output-complete time and actual process exit.
• Validate the new telemetry path in a compiled binary, including healthy delivery, stalled headers/body/TLS, proxy/NO_PROXY, and native DNS stalls. Existing compiled HTTP tests are not a substitute for this acceptance.
• The local reproduction confirms a shared lifecycle defect, not the unique cause of the reporter's Windows delay. Node can also reproduce stalled telemetry; do not conclude that completed idle sockets universally prevent exit or that Bun is the sole cause.
• Node telemetry's private dispatcher does not inherit custom host-installed global dispatcher CA/lookup/proxy policies. Standard CLI proxy environment variables and a fixture CA via NODE_EXTRA_CA_CERTS were verified.
• New enabled-telemetry subprocess cases cover normal/error-free completion; SIGINT/EPIPE have existing runtime coverage but have not been directly combined with slow telemetry in the new process fixture.
• The separate npm registry configuration request in [bug]: bl lingers ~50 s after answering (Windows) #294 is outside this change.