emrg: turn lifecycle timer fix (TUI freeze on mid-turn auto-compact + GUI per-session timers)

[argszero]

Summary

Fixes two host-reported timer bugs with a single shared root cause: there was no authoritative turn lifecycle from the daemon, so clients derived timers from incomplete signals (send time, done frames, compact results).

Rant 2026-09-02T10:31:11 (TUI timer freeze): auto-compact triggered mid-turn (between tool-loop rounds) broadcasts a compact_result completed frame; the TUI treated it as turn end, killed the elapsed timer, cleared the status-bar timer, and left the terminal title stuck at [1:37] while the turn kept running.

Rant 2026-09-02T10:36:26 (GUI sidebar per-session timer): the sidebar needs a per-session elapsed timer for every running session (incl. background evolution), aligned to the actual turn start — queued requests must not count queue wait.

Changes

• daemon (emrg/server/daemon.py): authoritative turn_start (with real started_at) / turn_end broadcasts in the tool-loop wrapper, covering all turn sources (TUI/GUI/scheduler/evolution/upgrade — all run through this locked wrapper). turn_end is broadcast before _session_task_cwds is cleared so the frame stays cwd-filtered (wrong-cwd ghost connections never receive it, per rant 2026-08-25T17:38:56).
• TUI (emrg/client/app.py): mid-turn auto-compact no longer sets busy = False (the timer loop stays alive; only manual /compact when idle is treated as terminal); turn_start aligns _request_start to the real execution time; ESC interrupt now resets the terminal title (was leaking a stuck [m:ss]).
• GUI renderer: daemonBridge.ts tracks turnStartBySid (turn_start/turn_end handlers + cleanup on done/cancelled/disconnect); Sidebar.tsx renders a per-session elapsed timer for every running session, aligned to the authoritative turn start; Shell.tsx drives a 1s tick. New pure formatElapsed in sidebar.ts.
• Tests: daemon e2e test_turn_start_end_broadcast_lifecycle (two-client form covering start→output→end ordering and queue-boundary turns) + test_broadcast_cwd_filtered now also guards the turn frames; renderer +3 daemonBridge +4 sidebar tests.

Verification

• pytest: 1221 passed / 1 skipped (1222 collected, Agent.md synced)
• renderer: tsc clean, 494 passed (Agent.md breakdown synced: sidebar 16, daemonBridge 18)
• GUI main: 93 pass / 0 fail, node --check clean
• import check + CLI help green

Conflict awareness

PR #1104 (GUI history) also bumps Agent.md count lines (python 1221→1222, renderer 487→490). Whoever merges second must reconcile: combined target is pytest 1223 collected, renderer 497.

[argszero]

✅ LGTM — cycle cyc20260902-110446

Reviewed the full diff (daemon turn lifecycle broadcasts, TUI compact guard, GUI per-session timers) + verified locally on the branch: test_turn_start_end_broadcast_lifecycle + test_broadcast_cwd_filtered + doc-counts = 7 passed.

• daemon: turn_start (started_at = actual execution start) / turn_end broadcast from _run_tool_loop_locked covers all turn sources; turn_end is sent BEFORE _session_task_cwds.pop() so the frame stays cwd-filtered (ghost connections never see it) — correct fix for the test_broadcast_cwd_filtered regression.
• TUI: busy is set at send time (app.py:477/2011) so the turn_start alignment fires reliably; compact_result only resets state when idle (not busy) — mid-turn auto-compact no longer kills the timer; ESC title reset matches the done path.
• GUI: clearTurnTimer is idempotent across turn_end/done/cancelled/disconnect; Shell's 1s tick runs only while turnStartBySid is non-empty (no idle re-render cost); formatElapsed [m:ss] matches the TUI format.
• CI test + test-windows both pass (33585236110).

[argszero]

✅ LGTM — cycle cyc20260902-111336 (2nd)

Re-reviewed the updated branch (9fc3760 = merge of master after #1104 + Agent.md reconciliation: pytest 1223 collected, renderer 497, GUI main corrected to true static 100 incl. nav-policy). Full local verification on the merged branch: pytest 1222 passed/1 skipped, renderer tsc clean + 497 passed, GUI 93 pass/0 fail, import + CLI green. CI re-fired after the push: test + test-windows both pass (33586460986). No new issues introduced by the merge.

[argszero]

✅ LGTM — cycle cyc20260902-112723 (3rd consecutive)

Third independent-cycle approval. Head unchanged (9fc3760) since the previous review; CI test + test-windows both green (33586460986); mergeStateStatus CLEAN. Verified the final diff: daemon turn_start/turn_end broadcasts (turn_end before _session_task_cwds pop — cwd-filter preserved), TUI compact guard + ESC title reset + _request_start alignment, GUI turnStartBySid sidebar timers. Local verification on this head already passed in cyc111336 (pytest 1222/1 skip, renderer 497, GUI 93). Satisfies the 3-consecutive-LGTM gate — merging.