Problem

Claude ≥ 2.1.x writes a session-name slug on every transcript entry once a session gets named (auto or manual). comm_runtime/parsing_runtime/structured.py maps entry.slug → event subagent_name, so from that point on every main-chain event is mislabeled as subagent activity.

Observed failure cascade (live incident, Claude 2.1.220):

1. is_top_level_user_prompt rejects the job's own anchor user prompt → anchor_seen / prompt_activated never flip in the poll state machine.
2. _prompt_completion_is_eligible then gates off poll_exact_hook forever — the exact Stop-hook artifact (written by the provider hook with the job's own req_id) is never consumed.
3. The job zombie-runs after its terminal reply: active-task flag never cleared, plain ask refused (forced --chain/--silence), and a --chain result continuation queued behind the zombie parent is never delivered — the child result strands on disk.

Reproduced deterministically by replaying the affected session transcript through ClaudeLogReader.try_get_entries + _process_events: anchor missed before the patch, anchor_seen=True after.

Fix

• Parsing (root cause): drop slug from subagent markers in _event_record (it is the session name, not a subagent marker); propagate entry.isSidechain as event is_sidechain and honor it in is_top_level_user_prompt / assistant_identity so real sidechain filtering still works via the reliable field.
• Polling (recovery): add _orphaned_exact_hook fallback — when session-log anchor tracking failed, accept the exact hook artifact only with independent proof: req_id match (already enforced by load_event), event timestamp ≥ accepted_at, 180 s grace elapsed, hook session == tracked session, and an observably idle pane. This un-zombies stuck jobs and lets terminal replies clear the active-task flag.
• Chain delivery (no result loss): when a callback edge is already FAILED/TIMED_OUT at child completion, submit_callback_continuation used to return early and silently drop the child result. New deliver_orphaned_chain_result delivers it to the dispatching agent (edge.parent_agent) as a notice instead.

Verification

• Transcript replay regression: anchor detected, prompt activated, reply captured after patch.
• _orphaned_exact_hook boundary cases (7/7): busy pane, fresh event (< grace), pre-submission event, session mismatch, no pane proof, prompt-not-sent all rejected.
• deliver_orphaned_chain_result cases (8/8): notice recorded to dispatching agent with orphan diagnostics and full body; safe on missing bureau/parent.
• Existing suites on patched tree: test_claude_comm_parsing, test_claude_hook_results, test_claude_queued_prompt_activation, test_claude_execution_polling, test_v2_ccbd_dispatcher, test_v2_message_bureau_dispatcher_integration — 171 passed.
• Live deployment: daemon restart finalized the zombie job via the normal hook path, delivered the stranded chain continuation to the dispatching agent, edge reached done, and all agents returned to zero open jobs.