Two coupled fixes for the 2026-04-28 wedge where a half-closed S3 socket left download_to_file blocked inside iter_content for 7+ hours, coalescing seven sched:sync:plaud refs onto one frozen task. think/importers/plaud.py: - download_to_file: switch session.get(stream=True) to (connect=30s, read=45s) tuple timeout so per-chunk stalls fail fast. Wrap streaming in RequestException handling that returns False with a clear log. - Add throttled progress_cb (~5s) into iter_content; PlaudBackend.sync passes a closure that refreshes last_completed mid-download so the outer inactivity timer trips on per-step progress, not just on per-file completion. think/supervisor.py + think/scheduler.py + think/utils.py: - New TaskQueue.set_cap / enforce_deadlines: cmd_name-keyed runtime caps. supervise() ticks enforce_deadlines once per second; on cap breach it sends non-blocking SIGTERM (raw send_signal, not RunnerManagedProcess.terminate which blocks on wait(timeout=15)), records the termination start, then escalates to SIGKILL after 15s via a sidecar dict. Mirrors the existing _restart_requests pattern. - scheduler.load_config parses optional max_runtime per entry via new shared parse_duration_seconds helper (accepts int, Ns, Nm, Nh). Invalid values warn and drop the cap; the entry still loads. - scheduler.collect_runtime_caps feeds caps into the TaskQueue at supervisor boot, keyed on TaskQueue.get_command_name(cmd). Caps are off by default; absent max_runtime means no cap. Note: no historical sync:plaud import logs were reachable from this worktree to tune the schedule value. Operators can set "max_runtime": "30m" on the sync:plaud entry in journal/config/schedules.json — comfortably above normal sync duration, well below the 1h hourly cadence. Tests: new tests/test_plaud_download.py covers read-timeout failure, throttled progress callback, and the inactivity-timer trip when sync makes no progress. tests/test_utils.py covers parse_duration_seconds. tests/test_supervisor.py covers set_cap, SIGTERM on deadline, SIGKILL escalation after 15s, termination-state cleanup on ref exit, and the no-cap noop. tests/test_scheduler.py covers max_runtime parsing (valid string, valid int, invalid negative/garbage/wrong-type), and collect_runtime_caps filtering. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>