new kamil adc

python_app/orchestration/process_supervisor.py

@@ -23,6 +23,10 @@ logger = logging.getLogger(__name__)
 _LOG_MAX_BYTES = 8 * 1024 * 1024
 # Per-process force-kill deadline used on stop; each child gets its own window.
 _STOP_GRACE_SECONDS = 2.0
+# Cmdline marker for the Kamil ADC collector binary. The collector runs in its own
+# session, so it can outlive a force-killed acquisition producer while still holding
+# the L-Card E-502; we hunt it by name as a guaranteed device-release backstop.
+_KAMIL_COLLECTOR_MARKER = b"kamil_adc_collector"
 
 
 @dataclass(slots=True)
@@ -135,6 +139,9 @@ class ProcessSupervisor:
         """Start required pipeline binaries and wait until they are ready."""
         if self.is_running():
             raise RuntimeError("Acquisition processes are already running")
+        # Clean slate: an ADC collector orphaned by a prior crash/force-kill can
+        # still hold the E-502 and make a fresh acquisition fail to open.
+        self._kill_kamil_collectors()
 
         logger.info(
             "Starting pipeline from config %s (allow_clean_orchestrator_exit=%s)",
@@ -186,6 +193,7 @@ class ProcessSupervisor:
     def stop(self) -> None:
         """Stop acquisition-side processes, keep processor process intact."""
         self._stop_processes(["sweep_orchestrator", "data_preprocessor"])
+        self._kill_kamil_collectors()
 
     def stop_orchestrator(self) -> None:
         """Stop orchestrator process only."""
@@ -198,6 +206,7 @@ class ProcessSupervisor:
     def stop_all(self) -> None:
         """Stop all managed processes."""
         self._stop_processes(["sweep_orchestrator", "data_preprocessor", "data_processor"])
+        self._kill_kamil_collectors()
 
     def _spawn(self, name: str, command: list[str], *, allow_clean_exit: bool) -> None:
         """Spawn one process unless same process is already alive.
@@ -353,6 +362,32 @@ class ProcessSupervisor:
             except (ProcessLookupError, PermissionError):
                 pass
 
+    def _kill_kamil_collectors(self) -> None:
+        """SIGKILL any Kamil ADC collector found by name — a guaranteed device backstop.
+
+        The collector survives a force-killed producer (own session) and can hang
+        the E-502, so it is hunted by cmdline and its group killed hard, regardless
+        of who its parent is. Best-effort and never raises.
+        """
+        own_pid = os.getpid()
+        try:
+            pids = [int(entry.name) for entry in Path("/proc").iterdir() if entry.name.isdigit()]
+        except OSError:
+            return
+        killed = 0
+        for pid in pids:
+            if pid <= 1 or pid == own_pid:
+                continue
+            try:
+                cmdline = (Path("/proc") / str(pid) / "cmdline").read_bytes()
+            except OSError:
+                continue
+            if _KAMIL_COLLECTOR_MARKER in cmdline:
+                self._signal_group(pid, signal.SIGKILL)
+                killed += 1
+        if killed:
+            logger.warning("Force-killed %d orphaned Kamil ADC collector process(es)", killed)
+
     @staticmethod
     def _log_abnormal_stop_exit(process: ManagedProcess) -> None:
         """Log a process that self-exited abnormally around stop time.
@@ -523,6 +558,9 @@ class ProcessSupervisor:
             self._pidfile_path.write_text("", encoding="utf-8")
         except OSError:
             pass
+        # The ADC collector is a grandchild (spawned by the producer), not in the
+        # pidfile, so reap it separately by name.
+        self._kill_kamil_collectors()
 
     def _is_stale_pipeline_pid(self, pid: int) -> bool:
         """Return whether `pid` still runs one of our pipeline binaries/scripts.