Delegate TCL folding to server, fix bugs, and enhance stability

- Move TCL folding range computation from client to language server, avoiding duplicate regions.
- Serialize language server restarts to prevent multiple server instances running concurrently.
- Pin server-side Python dependencies to specific minor versions for improved stability and predictability.
- Make debug server connection non-fatal, allowing the language server to start even if the debugger isn't attached.
- Optimize semantic token generation by caching Tclint plugin commands.

server/pyproject.toml

@@ -4,9 +4,12 @@ version = "0.1.0"
 description = "Python language server for NX Postprocessor Support"
 requires-python = ">=3.8"
 dependencies = [
-    "pygls",
-    "packaging",
-    "tclint",
+    # Upper bounds guard against breaking API changes in bundled deps. The code
+    # is written against tclint 0.8.x (see tools/semantic_tokens.py); pin it so a
+    # re-bundle can't silently pull an incompatible major/minor.
+    "pygls>=1.3,<2",
+    "packaging>=24,<27",
+    "tclint>=0.8,<0.9",
 ]
 
 [dependency-groups]

server/src/_debug_server.py

@@ -27,7 +27,17 @@ if debugger_path:
 
     # 5678 is the default port, If you need to change it update it here
     # and in launch.json.
-    debugpy.connect(5678)
+    # Connecting requires the "Python debug server" listener (launch.json) to be
+    # up first. If it isn't (e.g. wrong launch config was used), don't crash the
+    # whole language server - just continue running without the debugger attached.
+    try:
+        debugpy.connect(5678)
+    except (ConnectionRefusedError, OSError) as exc:
+        print(
+            f"debugpy: could not connect to debug adapter on port 5678 "
+            f"({exc}); continuing without debugging.",
+            file=sys.stderr,
+        )
 
     # This will ensure that execution is paused as soon as the debugger
     # connects to VS Code. If you don't want to pause here comment this

server/src/tools/semantic_tokens.py

@@ -7,6 +7,23 @@ from common.load_data import standard_items
 import lsprotocol.types as lsp
 
 
+# Constructing a PluginManager scans entry points, and get_commands() rebuilds
+# the builtin command set on every call. Semantic tokens are requested often, so
+# cache the manager and the resolved commands per plugin set.
+_PLUGIN_MANAGER = None
+_COMMANDS_CACHE = {}
+
+
+def _load_commands(plugins):
+    global _PLUGIN_MANAGER
+    if _PLUGIN_MANAGER is None:
+        _PLUGIN_MANAGER = PluginManager()
+    key = tuple(plugins)
+    if key not in _COMMANDS_CACHE:
+        _COMMANDS_CACHE[key] = _PLUGIN_MANAGER.get_commands(list(plugins))
+    return _COMMANDS_CACHE[key]
+
+
 class TokenModifier(enum.IntFlag):
     deprecated = enum.auto()
     readonly = enum.auto()
@@ -47,7 +64,7 @@ TOKEN_TYPES = [
 
 class _Highlighter(Visitor):
     def __init__(self, plugins, custom_functions: dict[str : list[lsp.CompletionItem]]):
-        self._commands = PluginManager().get_commands(plugins)
+        self._commands = _load_commands(plugins)
         self._tokens = []
         self.custom_functions = custom_functions