fix(codex): surface hook blocks and empty turns instead of silent idle

[qiankunli]

codex core silently ends a turn when a UserPromptSubmit/SessionStart hook
blocks the prompt: turn/completed reports status=completed with no items,
no error event, and the prompt never enters the native thread history.
baton dropped hook/* notifications and treated the zero-output completion
as a normal end_turn, so the user saw their message swallowed
(incident bs_01KXCNW0WVA11NZH2F8FKTCJ5E, misread as "codex stuck").

• handle hook/completed: blocked/stopped userPromptSubmit/sessionStart
  hooks emit a warning _baton_notice with source path and block reason,
  and stash the reason on the active turn
• finishTurn: a completed driven turn with zero visible output emits an
  empty-turn warning notice, attributed to the hook block when known;
  interrupted turns and pure-compaction turns are exempt
• wire shapes verified against a live codex 0.144.1 app-server session
  (inject_items + turn/start repro also disproved the earlier theory that
  cross-provider context injection caused the empty turns)

Documents the normalization semantics introduced by the empty-turn fix:
codex hook/completed notifications (blocked/stopped userPromptSubmit or
sessionStart — the paths that make codex end a turn silently with the
prompt never entering native history) and zero-output completed turns
normalize to a warning _baton_notice; lifecycle is unchanged, terminal
state still flows through state_update. Claude has no wire counterpart
(in-process SDK hook failures surface on the error stream).