Harness driver waits for its child; adoption stops when nothing is pending; SDK calls time out

[navbytes]

Follow-ups from the roost hang investigation, written up in docs/2026-09-06-roost-hang-report.md (answers docs/2026-09-06-roost-hang-brief.md). Short version: the 2026-09-04 wedges were the pty driver exiting on its --timeout with roost still running under the closed pty master; H1/H2/H3 on the plugin side were measured and cleared.

Changes

• harness/pty-run.py never leaves its child behind. On timeout or --exit-when-done it SIGTERMs the child's process group, keeps draining for up to 3 s, SIGKILLs, reaps, and prints a timestamped stop line with the exit status. A pyte failure in the final snapshot no longer skips the capture files. (killpg on an exited, unreaped child raises EPERM on macOS, handled.)
• Adoption stops when nothing is pending. adoptNativeForks returns { adopted, adoptables }; retryAdopt stops on a genuine zero as well as on success, so the sibling loop of every fork batch no longer polls three times for nothing. A pass that threw reports -1, so a transient session.list failure still retries.
• 15 s timeout on SDK calls in the server half, the TUI entry point (session.list) and the transcript fetcher, via AbortSignal.timeout. A failed transcript fetch returns the deleted sentinel so the route retries instead of caching an empty transcript. src/tui/actions.ts is left untimed on purpose (session.prompt runs for a whole turn).
• USAGE: /ctree subcommands queue behind a running turn; /tree opens from the local journal.

Verification

• bun run typecheck clean; bun test 275 pass; bun run build clean.
• python3 harness/pty-run.py --timeout 2 -- sh -c 'trap "" HUP; sleep 100' returns in ~2 s with the grandchild gone.
• CTREE_E2E=1 bun test test/e2e: 16 pass.
• Reviewed; the review's blocking finding (a failed pass stopping the retry loop) is fixed and covered by a unit test.