fix(state): keep a blocked turn-end gate until the pane really works again [no-ticket]

[matthias-scale]

Symptom

On three hosts, every pane whose last turn ended with a blocking gate showed agent_status: working (or idle) in herdr agent list — never blocked — so the sidebar gate badge never appeared. The gate data itself was correct on the same record (tokens.closing_blocking = "1", state_labels.blocked = "gate", gates: [1 item]), which proves the report was accepted: handle_pane_report_agent only applies the gates array when the state report was accepted.

Confirmed cause

Live evidence on w1C:pK while blocked: screen_detection_skipped was absent (i.e. full_lifecycle_hook_authority_active() == false) while sibling non-gate panes had it true, and herdr agent explain w1C:pK returned state: working, rule: osc_title_working, evidence: "⠐ Prepare coaching kickoff notes with Davud" — a stale spinner OSC title. So the pane still detected claude (the authority was therefore effective), yet the full-lifecycle authority was no longer live, and the screen fallback won.

herdr:claude was ruled out as a competing authority: the installed integration hook only ever sends pane.report_agent_session, never a state report.

That leaves AppEvent::HookAuthorityRetired → retire_blocked_full_lifecycle_hook_authority_at. FullLifecycleHookOutputRetirement::observe armed on the first content change after the blocked report and then expired on wall clock alone, five seconds later, whether or not anything else was written. Ending a turn is itself a write: the agent repaints its prompt box as soon as the Stop hook returns, which bumps detection_content_seq (any non-empty PTY chunk does). So every gate was retired ~5s after it was raised, and the pane fell back to fallback_state.

Fix

• src/pane/agent_detection.rs — retirement now measures the window between the first and the latest content change. Continuing output (a resumed turn: tokens, spinner, counters) still retires the gate after the same grace; a single turn-end repaint followed by silence never does.
• src/terminal/state.rs — hook_authority_is_effective and the closing-block admission arm of route_full_lifecycle_hook_report no longer require the agent to be detected on screen. Absence of detection is not evidence of absence — at turn end the prompt box matches no rule, and the scan is skipped outright while the authority is live. Only a different detected agent or an observed process exit contradicts the hook; suppressed sources keep the stricter presence guard.

No protection is weakened: a stale frame still cannot resurrect a finished gate (visible_working_overrides_idle_hook / detected_working_overrides_idle_hook are unchanged and still guarded by authority.state == Idle), process exit and agent conflict still release the authority immediately.

Tests

New:

• a_turn_end_repaint_alone_does_not_retire_a_hook_gate — one repaint burst then silence never retires.
• a_blocked_closing_block_report_survives_unavailable_screen_detection — blocked report with detected_agent = None and a working fallback stays Blocked (failed before the fix: left: Working, right: Blocked).
• a_blocked_closing_block_report_ends_when_a_new_turn_retires_it — genuine retirement still returns the pane to Working.
• a_blocked_closing_block_report_yields_to_a_different_detected_agent, ..._yields_to_the_agent_process_exiting — the two real contradictions still win.

Updated: continuing_output_retires_a_hook_gate_after_the_grace_period and output_before_the_first_detection_tick_starts_the_grace_period now advance the content seq across the window, matching the new contract.

cargo test --bin herdr: 3456 passed, 3 failed — all pre-existing environment flakes unrelated to this diff (app::api::plugins::tests::manifest_action_invoke_injects_plugin_paths, server::headless::tests::render_and_stream_sends_terminal_frame_for_terminal_ansi_client, workspace::tests::generated_workspace_ids_are_short_base32_handles). cargo fmt --check and cargo clippy --bin herdr --all-features clean.

[matthias-scale]

No description provided.