feat(cli): explain stopped/unknown agents and tighten CLI diagnostics

Stopped and unknown agents
- kill/stop/list/term/config/send/transcript/resume report when a name
  stopped, why, by whom, and how to resume; unknown names get typo
  suggestions (tag-name aliases included). kill/stop of an already-stopped
  agent exits 0; stop flags a leftover process by session and suggests
  `hcom kill <pid>`.
- kill keeps hcom name semantics for leftover processes (base name or
  matching tag-name) and refuses to guess when a reused base name matches
  several; output says "Sent SIGTERM" / "had already exited" instead of
  process-group internals.

Lifecycle
- PTY exit cleanup finalizes through the atomic stop gate and reads the
  initiator recorded by `hcom kill`: one stopped event per stop, attributed
  to the real initiator instead of "pty".
- Stop events are timestamped after taking the write lock.

send
- Reject a lone bare word without '@' or '--' (was a silent broadcast).
- Reject self-only targets, including bigboss:DEVICE.
- Unmatched @targets get typo suggestions plus stopped-agent resume hints;
  empty thread deliveries say so instead of "no agents are active".

Other
- router: did-you-mean for unknown commands; `hcom 2 [--go] claud` reports
  an unknown tool.
- config: unknown keys are rejected (were read as "(not set)" and written
  to config.toml); env-only keys that are set stay readable; any dev_root
  spelling routes to its handler.
- transcript: out-of-range requests report the available range (also --json).
- events: --agent warns on never-seen names (not with --all/--wait);
  same-second events keep write order.
- resume of a live agent suggests fork.