fix(cli): renders on Windows stop cancelling when an old parent's pid is reused - #5358

Approve at 066eb11e. The walk now stops at a "parent" created after its child, and only a reused pid can look like that. Real parents are always created before their children, so the real chain is kept, and a terminal that really closes still cancels the render.

Checked

• The cause: on Windows the ancestor chain comes from one Win32_Process listing, following ParentProcessId. Windows doesn't reparent orphans, so a dead parent's pid can belong to an unrelated, newer process. The old walk followed it, and the render cancelled when that process exited.
• The guard is complete for reuse. A reused pid's new process can only start after the original parent died. That parent was alive when it created the child, so the new process always starts after the child, and parent.startedAt > child.startedAt catches it. The one exception is a clock step, which the body already covers under Limits.
• The time source: the listing uses CreationDate.ToFileTimeUtc(), which is an integer tick count. So /^\d+$/ plus BigInt parses it, and comparing BigInts avoids float precision loss.
• A missing or unparseable time leaves startedAt undefined, and the guard does nothing. That fails toward the old behaviour (watch the parent), never toward dropping a real one.
• Closing the terminal still cancels:
  • The truncated chain only loses entries above a reused pid, which were never real ancestors.
  • The shell and every real parent below it stay watched.
  • Closing the console also delivers SIGHUP, which aborts on its own.
• Scope: recordsToAncestors runs only for the off-Linux snapshot, so Linux is untouched. macOS reparents orphans to launchd, so a dead parent's pid isn't followed there.

Nits (non-blocking)

• The body says the macOS ps listing "carries no start time". lstart is a start time; it's only a string. The conclusion still holds, because macOS reparents orphans.
• Pre-existing, opposite direction: off Linux the watchdog only checks isAlive(pid), never identity (renderCancellation.ts:125-133). On Windows, if the real parent exits and its pid is reused within the 2 s poll, the render keeps running orphaned until that new process exits. The snapshot already records windows:<creation> identities, so comparing them in the off-Linux check would close this the same way Linux does.

Tests:

• orphanCleanup (all three files), renderCancellation.test.ts and processTree.test.ts pass 28/28.
• renderCancellation.process.test.ts fails 4 cases on this box, both at this head and at its base, so it's environmental; CI runs it.
• Mutants:
  • > → >= is caught by the same-tick test.
  • Dropping the guard is caught.
  • Never parsing startedAt is caught.
• Every check on this head passed before this review was posted.

— Rames