feat(claire): ✨ add session cleanup and kick delay logic

Co-Authored-By: Lilith Autocommit <noreply@atlilith.com>

.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"c5a10f83-d897-4eb1-aa4f-a4264850dbf3","pid":15738,"procStart":"Mon Jun  1 06:29:14 2026","acquiredAt":1780437249785}

src/claire/orchestrator/bootstrap.py

@@ -601,6 +601,17 @@ def ensure_running(
         and r.cwd.rstrip("/") == norm_target
     }
 
+    # Kill any pane lingering at this slug before spawning. Reaching here means
+    # the fast path above found no VALID live session to preserve, so a pane at
+    # the slug is an orphan from an earlier cycle whose discovery timed out
+    # (the tmux session spawned, but its uuid never got persisted). Left alone
+    # they accumulate — and since turns are delivered by slug match, a duplicate
+    # orchestrator would receive every turn too, splitting delivery. Best-effort.
+    try:
+        rcl.kill(match=slug, yes=True)
+    except RclaudeError:
+        pass
+
     # Spawn a fresh session. Pass `.mcp.json` via `--mcp-config`. Path must
     # be remote-absolute so the spawned `claude` reads it.
     try:
@@ -612,23 +623,37 @@ def ensure_running(
     except RclaudeError:
         return None
 
-    # Claude doesn't flush its session JSONL to disk until the first user
-    # message is processed — discovery would otherwise spin until timeout
-    # on an empty filesystem. Send a no-op kick so the JSONL appears.
-    # Wait briefly for tmux + claude to reach the prompt before sending.
-    time.sleep(2.0)
-    try:
-        _send_kick(rcl=rcl, cwd=effective_cwd)
-    except RclaudeError:
-        # Kick failure isn't fatal — discovery may still succeed if the
-        # session writes its JSONL for other reasons.
-        pass
-
-    new_uuid = discover_session(
-        cwd=effective_cwd, host=cfg.orchestrator.host, rclaude=rcl,
-        timeout_s=discover_timeout_s,
-        ignore_uuids=pre_uuids,
-    )
+    # Claude doesn't flush its session JSONL to disk until it processes its
+    # first message — until then discovery has nothing to find. The kick both
+    # introduces Claude to its role and triggers that first flush.
+    #
+    # Timing is a race: a freshly spawned `claude` needs a few seconds to
+    # reach a live input prompt, and a kick sent before then is silently
+    # dropped — the session never processes it, never flushes, discovery
+    # fails, and the orchestrator never comes up (observed: panes stuck at
+    # R:0 with the kick text unconsumed). The delay scales with host load, so
+    # a single fixed sleep is unreliable. Instead, poll: send the kick, look
+    # for the new session, and re-send on each miss until it appears or the
+    # budget is spent. Re-sending is harmless — Claude just re-acknowledges
+    # once ready. capture-pane can't gate readiness here: the orchestrator may
+    # live on a remote host, where capture-pane is unsafe.
+    new_uuid: str | None = None
+    time.sleep(2.0)  # let tmux attach + claude begin launching
+    deadline = time.time() + discover_timeout_s
+    while True:
+        try:
+            _send_kick(rcl=rcl, cwd=effective_cwd)
+        except RclaudeError:
+            pass  # transient — the next iteration re-sends
+        # Clamp to ≥1s so a tiny `discover_timeout_s` still gets one real
+        # discovery attempt; cap at 4s so a miss re-kicks promptly.
+        new_uuid = discover_session(
+            cwd=effective_cwd, host=cfg.orchestrator.host, rclaude=rcl,
+            timeout_s=max(1.0, min(4.0, deadline - time.time())),
+            ignore_uuids=pre_uuids,
+        )
+        if new_uuid is not None or time.time() >= deadline:
+            break
     if new_uuid is None:
         return None
     write_session_uuid(