Solving refresh bug which disables every tests

- term console via flatpak-spawn --host so host venvs resolve (bins.host_console_command)
- QSettings sync() before subprocess kill in choices/tested-refs dialogs
- console regression test: fails on the in-sandbox 0.2.1 console

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -27,6 +27,27 @@ Pre-built artifacts are published at
   runnable directly, no Python installation required on the host. Lua
   support still needs a system `lua` interpreter and the `lua-socket` /
   `lua-cjson` modules.
+* **AppImage** (`Testium-<version>-x86_64.AppImage`) — single-file
+  Linux binary, runnable directly:
+
+  ```sh
+  chmod +x Testium-*-x86_64.AppImage
+  ./Testium-*-x86_64.AppImage
+  ```
+
+  Requires `libfuse2` on the host (FUSE 2 — distinct from `fuse3`, which
+  most distros now ship by default):
+
+  | Distro | Package |
+  |--------|---------|
+  | Arch / CachyOS / Manjaro | `fuse2` |
+  | Debian trixie / Ubuntu 24.04+ | `libfuse2t64` |
+  | Debian bookworm / Ubuntu 22.04 | `libfuse2` |
+  | Fedora | `fuse-libs` |
+
+  If you can't install libfuse2 (e.g. minimal container), prefix the
+  invocation with `APPIMAGE_EXTRACT_AND_RUN=1` — the AppImage will
+  self-extract to `/tmp` on each run instead of FUSE-mounting.
 * **Flatpak bundle** (`testium.flatpak`) — install with:
 
   ```sh
@@ -41,17 +62,6 @@ Pre-built artifacts are published at
   `testium` command is available in the terminal (requires `~/.local/bin` in
   `PATH`, which most modern distributions provide by default).
 
-* **AppImage** (`Testium-<version>-x86_64.AppImage`) — a single self-contained
-  executable bundling its own Python. Make it executable and run it:
-
-  ```sh
-  chmod +x Testium-*-x86_64.AppImage
-  ./Testium-*-x86_64.AppImage -b mytest.tum
-  ```
-
-  As with the binary and Flatpak, `py_func` / `lua_func` items run under the
-  *host* Python / Lua so your own modules stay visible.
-
 Every channel ships the language server, so `testium lsp` (see
 [Editor support](#editor-support)) works out of the box from any of them.
 
@@ -119,6 +129,22 @@ A VSCode / VSCodium client extension (`testium_assist`) wraps `testium lsp`;
 the schema is built from testium itself, so new item types and parameters
 appear in the editor on the next testium upgrade with no client change.
 
+It is published on [Open VSX](https://open-vsx.org/extension/testium/testium-assist),
+so in **VSCodium, Cursor, Windsurf, Theia and code-server** it installs from the
+Extensions view (search `testium-assist`) or with
+`codium --install-extension testium.testium-assist`.
+
+**Microsoft VSCode** does not list Open VSX extensions, so install the `.vsix`
+by hand — download it from the Open VSX page above, then *Extensions → ⋯ →
+Install from VSIX…* or:
+
+```sh
+code --install-extension testium-assist-0.1.0.vsix
+```
+
+The extension runs `testium lsp`, so `testium` must be on the `PATH` (otherwise
+point the `testium.serverPath` setting at the binary/AppImage).
+
 ## Troubleshooting
 
 ### `wl_proxy_marshal_flags` symbol error

doc/manual/sphinx/source/modes.rst

@@ -23,3 +23,80 @@ graphical interface.
     :caption: call a test in batch mode
 
     testium -b test/my_test/main.tum
+
+.. _sec_language_server:
+
+Language server (editor support)
+--------------------------------
+
+*testium* ships a `Language Server Protocol
+<https://microsoft.github.io/language-server-protocol/>`_ server so that
+``.tum`` files get editor assistance — completion of test item types, hover
+documentation of their parameters, and an outline view — in any LSP-capable
+editor.
+
+The server speaks LSP over standard input/output and is started with:
+
+.. code-block:: text
+    :caption: start the language server
+
+    testium lsp
+
+It is not meant to be launched directly by the user: an editor's LSP client
+spawns it and drives the exchange. A VSCode / VSCodium client extension,
+*testium_assist*, is provided for that purpose; any other LSP client (Neovim,
+Emacs ``lsp-mode``, …) can be pointed at ``testium lsp`` as well.
+
+The information the server exposes is the test item schema, which can also be
+dumped as JSON for inspection or tooling:
+
+.. code-block:: text
+    :caption: dump the item / parameter schema
+
+    testium schema
+
+Because the schema is built from *testium* itself, every new item type or
+parameter becomes available in the editor on the next *testium* upgrade, with
+no change to the client.
+
+The language server is included in the pre-built binary, Flatpak and AppImage
+releases. For a source or wheel installation, pull the optional ``lsp``
+dependencies:
+
+.. code-block:: text
+    :caption: enable the language server for a wheel / source install
+
+    pip install 'testium[lsp]'
+
+Installing the VSCode / VSCodium extension
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The *testium_assist* client extension is published on `Open VSX
+<https://open-vsx.org/extension/testium/testium-assist>`_, the registry used by
+VSCodium, Cursor, Windsurf, Eclipse Theia and code-server. In those editors,
+open the Extensions view and search ``testium-assist``, or install it from the
+command line:
+
+.. code-block:: text
+    :caption: install in VSCodium and other Open VSX editors
+
+    codium --install-extension testium.testium-assist
+
+Microsoft *VSCode* uses a different marketplace that does not list Open VSX
+extensions, so install the packaged ``.vsix`` by hand. Download it from the
+Open VSX page linked above, then either choose *Extensions* → *⋯* →
+*Install from VSIX…* in the UI, or run:
+
+.. code-block:: text
+    :caption: install the .vsix in Microsoft VSCode
+
+    code --install-extension testium-assist-0.1.0.vsix
+
+The extension launches ``testium lsp``, so the ``testium`` command must be on
+the ``PATH``. If *testium* is installed elsewhere — a specific binary or an
+AppImage — point the ``testium.serverPath`` setting at it instead.
+
+Once installed, open a ``.tum`` file: completion of item types, hover
+documentation and the outline view become available. If nothing happens, check
+that no ``files.associations`` entry forces ``*.tum`` to another language (it
+must stay the ``tum`` language the extension provides).

doc/manual/sphinx/source/test_items/let_test_item.rst

@@ -9,9 +9,9 @@ This element is of the following form:
     - let:
         name: Let Item
         values:
-            key1: value1
-            key2: value2
-            key3: <| $(variable)[$(loop_index)] |>
+            - key1: value1
+            - key2: value2
+            - key3: <| $(variable)[$(loop_index)] |>
 
 The ``let`` element is used to set values in the global directory.
 

doc/manual/sphinx/source/tum_syntax.rst

@@ -51,8 +51,8 @@ The parameter file can be specified in the `.tum` file root:
     :caption: configuration files definition in the main `.tum` test file
 
     config_file:
-        config1.yaml
-        config2.yaml
+        - config1.yaml
+        - config2.yaml
 
     main:
         name: Test example
@@ -232,6 +232,15 @@ list of the main test item (and eventually of the loop test item).
 TUM file ``main`` item is itself a variant of test items with a name and an
 step list attributes.
 
+.. note::
+
+   Each test item declares the parameters it accepts. When a ``.tum`` file
+   uses a key the item does not know, *testium* emits a warning listing the
+   accepted parameter names (catching typos such as ``param_filee`` for
+   ``param_file``); a missing **required** parameter aborts loading with an
+   error pointing at the source ``.tum`` file. Valid existing tests are
+   unaffected.
+
 
 
 .. toctree::

release_note.txt

@@ -1,3 +1,22 @@
+version 0.2.2
+==============
+- Flatpak sandbox issue fixed for term console. Now a term console is
+  exactly like a host console.
+- Persistence fix of dialogs in case of flatpak.
+
+version 0.2.1
+==============
+- Faster test loading, especially for large tests built from jinja
+  templates and ``!include``: compiled jinja templates are cached and
+  reused (a file included many times is compiled once), rendering happens
+  in memory instead of through a temporary file, and YAML is parsed with
+  the libyaml C loader when available. Typical load time is 3-6x lower on
+  include / template-heavy tests; behaviour is unchanged.
+- Fix: a nested list holding more than one step under ``steps`` no longer
+  duplicates its entries while the step tree is built.
+- New load-time benchmark under ``test/benchmark/`` (synthetic-tree
+  generator + in-process timing harness) to measure the load pipeline.
+
 version 0.2
 ==============
 - Test items: each item type now declares its accepted parameters

src/VERSION

@@ -1 +1 @@
-0.2
+0.2.2

src/testium/api/termconsole.py

@@ -81,9 +81,13 @@ class TermConsole(Console):
                                    bufsize=0)
 
         else:
-            self.term = pexpect.spawn(  shell_cmd,
-                                        echo=False,
-                                        cwd=self.ppath)
+            # In Flatpak this returns a `flatpak-spawn --host` wrapper so the
+            # console behaves like a host shell (matching py_func / lua_func /
+            # run); elsewhere it's the chosen command unchanged.
+            from interpreter.utils import bins
+            argv = bins.host_console_command(shell_cmd, self.ppath)
+            self.term = pexpect.spawn(argv[0], args=argv[1:],
+                                      echo=False, cwd=self.ppath)
 
         self.q = BytesStore()
         self.t = threading.Thread(target=self.enqueue_output)

src/testium/interpreter/test_items/dialog_choices_files/choices_dialog.py

@@ -221,6 +221,11 @@ def main(args, conn=None):
 
     if conn:
         settings.setValue(SettingsLastChoices, result)
+        # Flush before sending: the parent terminates this subprocess as soon
+        # as it reads the result, so the QSettings destructor never runs and
+        # the write would race the kill (lost under Flatpak — see the
+        # tested-references dialog for the full rationale).
+        settings.sync()
         conn.send([result, success])
         conn.close()
     else:

src/testium/interpreter/test_items/test_item_console.py

@@ -344,7 +344,7 @@ class TestItemConsoleReadUntil(TestItemConsoleAction):
     def execute(self):
         cons = self.get_console()
         ru = self._prms.expanse(self._read_until)
-        read_timeout = int(self._prms.getParam("timeout", default=-1, processed=True))
+        read_timeout = float(self._prms.getParam("timeout", default=-1, processed=True))
         mute = self._prms.getParam("mute", default=False, processed=True)
         if read_timeout < 0:
             read_timeout = None

src/testium/interpreter/test_items/tested_references_files/tested_refs_dialog.py

@@ -76,6 +76,12 @@ def main(args, conn=None):
 
     if conn:
         settings.setValue(SettingsLastReference, result)
+        # Flush to disk *before* handing the result back: as soon as the parent
+        # receives it on the pipe it terminates this subprocess (SIGTERM, no
+        # handler), so the QSettings destructor never runs. Without sync() the
+        # write races the kill and is lost — reliably so under Flatpak, where
+        # the .conf is atomically renamed on the slower ~/.var/app overlay.
+        settings.sync()
         conn.send([result, success])
         conn.close()
     else:

src/testium/interpreter/test_set.py

@@ -29,6 +29,51 @@ def _build_item_path(item) -> str:
     return " > ".join(reversed(parts))
 
 
+def _flatten_actions(actions, out, parent_seq_name):
+    """Expand nested lists and included ``sequence`` entries into ``out`` as a
+    flat list of single test-item dicts, propagating each sequence's source
+    filename onto its items.
+
+    Replaces the previous approach, which spliced each entry into the step
+    list and rebuilt the whole list every time (O(n^2) over the step list, and
+    a rebuild that duplicated entries when a nested list held more than one
+    element). This single forward pass is linear.
+    """
+    for idx, action in enumerate(actions):
+        # a bare list raises its elements to the same level
+        if isinstance(action, (list, tuple)):
+            _flatten_actions(action, out, parent_seq_name)
+            continue
+        # a NoneType (e.g. pointing at an unused alias) contributes nothing
+        if action is None:
+            continue
+        # a 'sequence' (an included file) is spliced in, with its filename
+        # propagated onto each of its items
+        if isinstance(action, dict) and "sequence" in action:
+            sequence = action["sequence"]["data"]
+            f = action["sequence"]["filename"]
+            if isinstance(sequence, dict):
+                sequence = [{k: v} for k, v in sequence.items()]
+            # Case of an empty sequence
+            elif sequence is None:
+                tm.print_info(
+                    f"An empty sequence is loaded in '{parent_seq_name}'."
+                )
+                sequence = []
+            elif not isinstance(sequence, list):
+                raise ETUMSyntaxError(
+                    f"Syntax error in '{parent_seq_name}' step number {idx+1}. Sequence definition: '{str(action)}'",
+                    f
+                )
+            for s in sequence:
+                if isinstance(s, dict) and s:
+                    s[list(s.keys())[0]]["seq_filename"] = f
+            _flatten_actions(sequence, out, parent_seq_name)
+            continue
+
+        out.append(action)
+
+
 class TestSet:
     def __init__(
         self,
@@ -434,56 +479,16 @@ class TestSet:
                 f"No valid list of actions in sequence {parent_seq_name}",
                 file_name
             )
-        # first we merged to the same level 'sequence dict entries and list within the list
-        counter = 0
         test_dir = tm.gd("test_directory")
-        la = len(parent_seq_actions)
-        while counter < la:
-            action = parent_seq_actions[counter]
-            # if action is a list raise up to the the same level,
-            # ie insert action element into the parent_seq_actions
-            if isinstance(action, (list, tuple)):
-                parent_seq_actions[counter : counter + 1] = action
-                parent_seq_actions = (
-                    parent_seq_actions[:counter]
-                    + action
-                    + parent_seq_actions[counter + 1 :]
-                )
-                la = len(parent_seq_actions)
-                continue
-            # if action is a NoneType skip and continue
-            # (when pointing to an unused alias for instance)
-            if action is None:
-                counter += 1
-                continue
-            # if action is a sequence we insert its entry into the action list
-            if "sequence" in action:
-                sequence = action["sequence"]["data"]
-                f = action["sequence"]["filename"]
-                if isinstance(sequence, dict):
-                    sequence = [{k: v} for k, v in sequence.items()]
-                # Case of an empty sequence
-                elif sequence is None:
-                    tm.print_info(
-                        f"An empty sequence is loaded in '{parent_seq_name}'."
-                    )
-                    sequence = []
-                elif not isinstance(sequence, list):
-                    raise ETUMSyntaxError(
-                        f"Syntax error in '{parent_seq_name}' step number {counter+1}. Sequence definition: '{str(action)}'",
-                        f
-                    )
-                for s in sequence:
-                    s[list(s.keys())[0]]["seq_filename"] = f
-                parent_seq_actions = (
-                    parent_seq_actions[:counter]
-                    + sequence
-                    + parent_seq_actions[counter + 1 :]
-                )
-                la = len(parent_seq_actions)
-                continue
 
-            # Action is now for sure a list of dict of length 1
+        # Flatten nested lists and included 'sequence' entries to the same level
+        # in one linear pass (was an in-place splice + full list rebuild per
+        # entry: O(n^2) over the step list).
+        flat_actions = []
+        _flatten_actions(parent_seq_actions, flat_actions, parent_seq_name)
+
+        for action in flat_actions:
+            # Action is now for sure a dict of length 1
             k = list(action.keys())[0]
             if action[k].get("seq_filename", None) is None:
                 action[k]["seq_filename"] = file_name
@@ -546,8 +551,6 @@ class TestSet:
                     action[k]["seq_filename"]
                 )
 
-            counter += 1
-
         return ret
 
     def tree(self):

src/testium/interpreter/utils/bins.py

@@ -19,6 +19,7 @@ Public API
 
 import atexit
 import os
+import shlex
 import shutil
 import subprocess
 import tempfile
@@ -177,6 +178,27 @@ def flatpak_host_spawn(interp_bin, cmd_args, host_cwd, extra_env=None):
     return spawn
 
 
+def host_console_command(shell_cmd, cwd):
+    """Build the argv to start *shell_cmd* as an ordinary interactive console.
+
+    *shell_cmd* is the command the caller chose (a string — shell-split — or
+    an argv list); the choice is preserved verbatim.
+
+    Outside Flatpak the command is returned unchanged. Inside Flatpak a bare
+    spawn would run in the sandbox under the runtime python3, so a host venv
+    (``/path/venv/bin/python3 -m mod``) can't see its pip deps. We simply run
+    it on the host with ``flatpak-spawn --host`` so it behaves like any other
+    terminal: flatpak-spawn passes the current environment through unchanged
+    and the shell (sourced venv, profile, …) sets things up as the user wants.
+    No env forwarding or scrubbing — the launcher's leaked PYTHONPATH points at
+    /app paths absent on the host, so it's inert there.
+    """
+    argv = shlex.split(shell_cmd) if isinstance(shell_cmd, str) else list(shell_cmd)
+    if not _in_flatpak():
+        return argv
+    return ["flatpak-spawn", "--host", f"--directory={cwd}", *argv]
+
+
 def _which_host_flatpak(name):
     """Resolve a binary name (or absolute path) on the host via flatpak-spawn.
 

src/testium/interpreter/utils/include.py

@@ -6,10 +6,10 @@ from runtime.tum_except import ETUMFileError
 from interpreter.utils.template import template_to_test
 from copy import copy
 from interpreter.utils.globdict import global_dict
-from interpreter.utils.yaml_load import yaml_load
+from interpreter.utils.yaml_load import yaml_load, YAML_BASE_LOADER
 
 
-class TUMLoaderNoIncludes(yaml.Loader):
+class TUMLoaderNoIncludes(YAML_BASE_LOADER):
 
     def __init__(self, stream):
 

src/testium/interpreter/utils/template.py

@@ -1,33 +1,74 @@
+import io
 import os
 from sys import exc_info
-from jinja2 import Template
+from jinja2 import Environment
 from jinja2.exceptions import TemplateSyntaxError, TemplateError, UndefinedError
-from tempfile import TemporaryFile
 from interpreter.utils.yaml_load import print_yaml
 from runtime.tum_except import ETUMSyntaxError
 
 
+# One Environment reused for every render (default settings, i.e. identical
+# behaviour to jinja2.Template), plus a compiled-template cache so a file that
+# is included many times — or a test that is reloaded — is compiled only once.
+# Jinja compilation is the expensive step; render (variable substitution) stays
+# per-call. Cache is keyed on path + mtime + size so an edited file recompiles.
+_ENV = Environment()
+_template_cache = {}  # abspath -> (mtime_ns, size, compiled_template)
+
+
+class _RenderedStream(io.StringIO):
+    """A rendered template kept in memory.
+
+    Carries ``root`` (and ``name``) so the YAML loader resolves ``!include``
+    paths exactly as it did from the on-disk temp file this replaces — without
+    the write + seek + read round-trip (one temp file per included file). That
+    round-trip is pure overhead, and especially costly on slow storage.
+    """
+
+
+def _compiled_template(filename: str):
+    """Return the compiled jinja template for *filename*, reusing the cached
+    one when the file is unchanged (path + mtime + size)."""
+    key = os.path.abspath(filename)
+    try:
+        st = os.stat(filename)
+    except OSError:
+        st = None
+    if st is not None:
+        cached = _template_cache.get(key)
+        if (cached is not None
+                and cached[0] == st.st_mtime_ns
+                and cached[1] == st.st_size):
+            return cached[2]
+    with open(filename, "r") as f:
+        source = f.read()
+    template = _ENV.from_string(source)  # compile (may raise TemplateSyntaxError)
+    if st is not None:
+        _template_cache[key] = (st.st_mtime_ns, st.st_size, template)
+    return template
+
+
 def template_to_test(filename: str, params: list):
     """ Function which processes an eventual jinja2 template to a test file
     """
-    # Temporary file created to receive the processed include
-    # file
-    tmpf = TemporaryFile('w+t')
-    with open(filename, 'r') as f:
-        try:
-            j2_template = Template(f.read())
-        except TemplateError as e:
+    # Compile (cached) — a syntax error in the template surfaces here.
+    try:
+        j2_template = _compiled_template(filename)
+    except TemplateError as e:
+        with open(filename, "r") as f:
             print_yaml(f, filename)
-            type, value, tb = exc_info()
-            msg = "Template error"
-            if hasattr(value, 'lineno'):
-                msg = msg + f" on line {value.lineno}: "
-            else:
-                msg += ": "
-            raise ETUMSyntaxError(msg + str(e), filename)
+        type, value, tb = exc_info()
+        msg = "Template error"
+        if hasattr(value, 'lineno'):
+            msg = msg + f" on line {value.lineno}: "
+        else:
+            msg += ": "
+        raise ETUMSyntaxError(msg + str(e), filename)
+
+    # Render into memory (no temp file).
     try:
         params["include_directory"] = os.path.dirname(os.path.abspath(filename))
-        tmpf.write(j2_template.render(params))
+        rendered = j2_template.render(params)
     except TemplateSyntaxError as e:
         raise ETUMSyntaxError(f"""Template loading of file '{filename}' with following parameters '{str(params)}'
 Syntax error in template: {e.message}""")
@@ -42,8 +83,7 @@ Template rendering error: {e.message}""")
         raise ETUMSyntaxError(f"""Template loading of file '{filename}' with following parameters '{str(params)}'
 Unexpected error: {str(e)}""")
 
-    # return to begining of the temp file
-    tmpf.seek(0, os.SEEK_SET)
-    tmpf.root = os.path.dirname(filename)
-
-    return tmpf
+    stream = _RenderedStream(rendered)
+    stream.root = os.path.dirname(filename)
+    stream.name = filename
+    return stream

src/testium/interpreter/utils/test_init.py

@@ -11,7 +11,7 @@ import api.testium as tm
 import interpreter.utils.globdict as globdict
 import interpreter.utils.settings as prefs
 from interpreter.utils.paths import testium_path
-from interpreter.utils.yaml_load import yaml_load
+from interpreter.utils.yaml_load import yaml_load, YAML_BASE_LOADER
 from interpreter.utils import clear_recursively
 from runtime.tum_except import ETUMSyntaxError
 from interpreter.utils.params import expanse, eval_func_init
@@ -89,7 +89,7 @@ def locate_report_file(rep_file):
 def yamltodict(param_file, silent=True):
     # load of the file
     with open(param_file, "r") as fd:
-        dp = yaml_load(fd, param_file, yaml.Loader)
+        dp = yaml_load(fd, param_file, YAML_BASE_LOADER)
 
     if dp is None:
         tm.print_info(f"The YAML file '{param_file}' is empty.")

src/testium/interpreter/utils/version.py

@@ -1,10 +1,54 @@
+import atexit
 import os
+import stat
 import sys
+import tempfile
 from importlib import import_module
 
 import interpreter.utils.settings as prefs
 import api.testium as tm
 
+
+# When running inside a Flatpak, the host /usr/bin/git is reachable at
+# /run/host/usr/bin/git but linked against host glibc/zlib, which the
+# sandbox can't load (``libz-ng.so.2`` not found). gitpython resolves git
+# eagerly on import and would crash the whole test run. We install a
+# tiny shell wrapper under /tmp that forwards to ``flatpak-spawn --host
+# git``, and point gitpython at it via ``GIT_PYTHON_GIT_EXECUTABLE``.
+_HOST_GIT_WRAPPER = None
+
+
+def _setup_flatpak_git():
+    global _HOST_GIT_WRAPPER
+    if not os.path.isfile("/.flatpak-info"):
+        return
+    if _HOST_GIT_WRAPPER is not None:
+        return
+    fd, path = tempfile.mkstemp(prefix="testium-git-host-", suffix=".sh", dir="/tmp")
+    with os.fdopen(fd, "w") as f:
+        f.write('#!/bin/sh\nexec flatpak-spawn --host git "$@"\n')
+    os.chmod(path, stat.S_IRWXU)
+    _HOST_GIT_WRAPPER = path
+    atexit.register(_cleanup_flatpak_git)
+    os.environ["GIT_PYTHON_GIT_EXECUTABLE"] = path
+    # Silence gitpython's warning if its refresh probe ever still fails;
+    # the wrapper itself should make the probe succeed.
+    os.environ.setdefault("GIT_PYTHON_REFRESH", "quiet")
+
+
+def _cleanup_flatpak_git():
+    global _HOST_GIT_WRAPPER
+    if _HOST_GIT_WRAPPER and os.path.isfile(_HOST_GIT_WRAPPER):
+        try:
+            os.unlink(_HOST_GIT_WRAPPER)
+        except OSError:
+            pass
+        _HOST_GIT_WRAPPER = None
+
+
+_setup_flatpak_git()
+
+
 _cached_versions = {}
 
 def repo_rev(path):

src/testium/interpreter/utils/yaml_load.py

@@ -1,3 +1,4 @@
+import yaml
 from yaml.parser import ParserError
 from yaml import load, Loader
 from yaml.scanner import ScannerError
@@ -5,6 +6,12 @@ from api.testium import print_debug
 from runtime.tum_except import ETUMSyntaxError
 import io
 
+# Use the libyaml-backed loader (much faster parsing) when PyYAML was built
+# with it, falling back to the pure-Python loader otherwise. The C loader
+# raises the same ParserError/ScannerError and supports the same custom
+# constructors (!include) and construct_* helpers the TUM loaders rely on.
+YAML_BASE_LOADER = yaml.CLoader if getattr(yaml, "__with_libyaml__", False) else yaml.Loader
+
 
 def print_yaml(file: io.TextIOWrapper, file_name):
     """ Prints YAML file if debug mode is activated.
@@ -21,10 +28,10 @@ def yaml_load(file, real_file_name: str, loader: Loader):
         return load(file, loader)
 
     except ParserError as e:
-        if isinstance(file, io.TextIOWrapper):
+        if isinstance(file, (io.TextIOWrapper, io.StringIO)):
             print_yaml(file, real_file_name)
         raise ETUMSyntaxError(f"yaml file parsing error: " + str(e), real_file_name)
     except ScannerError as e:
-        if isinstance(file, io.TextIOWrapper):
+        if isinstance(file, (io.TextIOWrapper, io.StringIO)):
             print_yaml(file, real_file_name)
         raise ETUMSyntaxError("yaml file scanning error: " + str(e), real_file_name)

src/testium/main_win/desktop_integration.py

@@ -0,0 +1,95 @@
+"""Install a desktop entry + icon under the user's data dir so desktop shells
+show the testium icon in the task bar / dock.
+
+On a native Wayland session GNOME takes a window's task-bar icon from the
+``.desktop`` file whose name (or ``StartupWMClass``) matches the window
+``app_id`` — ``QGuiApplication.setWindowIcon`` is ignored there. The portable
+channels (source checkout, PyInstaller binary, AppImage) install no system
+desktop file, so we drop an idempotent one in ``~/.local/share``. The window
+``app_id`` is set to ``testium`` (see ``QApplication.setDesktopFileName`` in
+``testium_win``), which is exactly this file's base name.
+
+Flatpak ships its own ``org.testium.Testium.desktop`` and keeps its own app id,
+so the caller skips this integration there.
+"""
+import os
+import sys
+
+from PySide6.QtCore import Qt
+from PySide6.QtGui import QPixmap
+
+# Must match QApplication.setDesktopFileName(...) for the GUI, and is used as
+# both the desktop-file base name and the StartupWMClass.
+APP_ID = "testium"
+
+
+def _launch_command():
+    """Best-effort Exec= for the menu entry. Not needed for icon matching, but
+    makes the entry actually launchable when possible."""
+    appimage = os.environ.get("APPIMAGE")
+    if appimage:
+        return f'"{appimage}"'
+    if getattr(sys, "frozen", False):
+        return f'"{os.path.abspath(sys.executable)}"'
+    argv0 = os.path.abspath(sys.argv[0]) if sys.argv and sys.argv[0] else ""
+    if argv0 and os.path.exists(argv0):
+        return f'"{os.path.abspath(sys.executable)}" "{argv0}"'
+    return f'"{os.path.abspath(sys.executable)}" -m testium'
+
+
+def ensure_desktop_entry():
+    """Create (or refresh) ~/.local/share icon + desktop entry. Best-effort:
+    any failure is swallowed so it can never take the GUI down.
+
+    Freedesktop-only: a no-op off Linux (Windows / macOS use the window icon)."""
+    if not sys.platform.startswith("linux"):
+        return
+    try:
+        data_home = os.environ.get("XDG_DATA_HOME") or os.path.join(
+            os.path.expanduser("~"), ".local", "share"
+        )
+        icon_dir = os.path.join(data_home, "icons", "hicolor", "256x256", "apps")
+        app_dir = os.path.join(data_home, "applications")
+        icon_path = os.path.join(icon_dir, f"{APP_ID}.png")
+        desktop_path = os.path.join(app_dir, f"{APP_ID}.desktop")
+
+        os.makedirs(icon_dir, exist_ok=True)
+        os.makedirs(app_dir, exist_ok=True)
+
+        # Icon: render the bundled Qt resource to a PNG once. Requires a live
+        # QGuiApplication (the caller creates it before calling us).
+        if not os.path.isfile(icon_path):
+            pixmap = QPixmap(u":/black/testium_logo.png")
+            if not pixmap.isNull():
+                pixmap = pixmap.scaled(
+                    256, 256,
+                    Qt.AspectRatioMode.KeepAspectRatio,
+                    Qt.TransformationMode.SmoothTransformation,
+                )
+                pixmap.save(icon_path, "PNG")
+
+        # Absolute Icon= path so the shell resolves it without an icon-cache
+        # refresh; StartupWMClass lets X11 / XWayland match too.
+        desktop = (
+            "[Desktop Entry]\n"
+            "Type=Application\n"
+            "Name=Testium\n"
+            "Comment=Test sequencer\n"
+            f"Icon={icon_path}\n"
+            f"Exec={_launch_command()} %f\n"
+            "Terminal=false\n"
+            f"StartupWMClass={APP_ID}\n"
+            "Categories=Utility;Development;\n"
+        )
+
+        # Write only when missing or changed, to avoid needless menu churn.
+        current = None
+        if os.path.isfile(desktop_path):
+            with open(desktop_path, "r") as fh:
+                current = fh.read()
+        if current != desktop:
+            with open(desktop_path, "w") as fh:
+                fh.write(desktop)
+    except Exception:
+        # Desktop integration is a nicety, never a hard requirement.
+        pass

src/testium/main_win/test_file_manager.py

@@ -51,14 +51,18 @@ class TestFileManager:
         w.disconnect_signals()
         # Snapshot user-selected checkboxes and fold state so they survive a
         # reload of the same file (same logic as session-restore through prefs).
+        # checkList works only if show_checkboxes is True
         previous_check_list = w.treeTests.getCheckList()
         previous_fold_list = w.treeTests.getFoldList()
         previous_count = w.treeTests.getItemCount()
         self.clear_process()
-        if self.load(file_name) and w.test_service is not None:
-            if w.treeTests.getItemCount() == previous_count:
-                w.treeTests.restoreCheckList(previous_check_list, w.test_service)
+        if self.load(file_name) and \
+            w.test_service is not None and \
+            w.treeTests.getItemCount() == previous_count:
+                if prefs.settings.show_checkboxes :
+                    w.treeTests.restoreCheckList(previous_check_list, w.test_service)
                 w.treeTests.restoreFoldList(previous_fold_list)
+
         w.reconnect_signals()
 
     def _make_progress(self, w):

src/testium/main_win/testium_win.py

@@ -678,6 +678,24 @@ def MainWin(
     debug=False,
 ):
     app = QApplication(sys.argv)
+    # Application identity so desktop shells (GNOME, ...) show the testium
+    # icon in the task bar / dock instead of a generic one. On Wayland this
+    # sets the surface app_id; on X11/XWayland it sets WM_CLASS, so the window
+    # stops inheriting the launcher's class (e.g. "python3" under the AppImage,
+    # which is what GNOME was keying the wrong icon off) and the window icon
+    # below is used as the fallback. In Flatpak the id must be the Flatpak app
+    # id so it matches the installed desktop file.
+    app.setApplicationName("Testium")
+    app.setApplicationDisplayName("Testium")
+    app.setDesktopFileName(os.environ.get("FLATPAK_ID", "testium"))
+    app.setWindowIcon(QIcon(u":/black/testium_logo.png"))
+    # On native Wayland the task-bar icon comes from an installed desktop file
+    # matched to the app_id, not from setWindowIcon(). Flatpak ships its own;
+    # for the other Linux channels drop an idempotent one under ~/.local/share.
+    # Windows / macOS use the window icon set above, so this is Linux-only.
+    if sys.platform.startswith("linux") and not os.environ.get("FLATPAK_ID"):
+        from main_win.desktop_integration import ensure_desktop_entry
+        ensure_desktop_entry()
     ui = MainWindow(
         test_file,
         config_files,

test/benchmark/.gitignore

@@ -0,0 +1 @@
+cases/

test/benchmark/README.md

@@ -0,0 +1,116 @@
+# Load-time benchmark
+
+Measures how long *testium* takes to **load** a `.tum` test tree — template
+rendering (jinja) + YAML parsing + test-tree construction — *without* executing
+it. Purpose: get reproducible numbers before/after load-path optimisations, and
+attribute any gain to a specific part of the pipeline.
+
+It is meant for *very long* tests, the kind you can build with `jinja` loops and
+`!include`, where load time becomes noticeable.
+
+## Files
+
+| File | Role |
+|------|------|
+| `gen_bench_test.py` | Generates a synthetic `.tum` tree (the test input). |
+| `load_bench.py` | Drives the **real** loader in-process and times it. |
+| `run.sh` | Convenience: generate + time across profiles, using the project venv. |
+| `cases/` | Generated trees (git-ignored, recreated on demand). |
+
+The benchmark `.tum` files are **generated**, not committed — the generator is
+the artifact. They use only `let` leaves and `group` containers, so loading has
+no runtime side effect (no subprocess, no `<| |>` eval) and the timing reflects
+the parse/build pipeline alone.
+
+## Quick start
+
+```bash
+# default matrix (all profiles), 5 repeats each
+./test/benchmark/run.sh
+
+# one profile at one size
+./test/benchmark/run.sh repeat 2000
+
+# more repeats for a tighter min
+REPEAT=10 ./test/benchmark/run.sh includes 1000
+```
+
+`run.sh` uses the project venv at `test/tmp/.venv` (created by `./run.sh`). If it
+is missing, run `./run.sh` once first.
+
+To drive the harness directly on any `.tum` (not just generated ones):
+
+```bash
+test/tmp/.venv/bin/python3 test/benchmark/load_bench.py --repeat 5 --quiet path/to/main.tum
+```
+
+## Profiles
+
+Each profile isolates one cost. `--size` is the profile-specific count.
+
+| Profile | What it builds | Stresses |
+|---------|----------------|----------|
+| `flat` | one main file, N inline `let` steps | big YAML parse + linear object build |
+| `includes` | main `!include`s N **distinct** sub-files | per-include template+YAML+tempfile, `sequence` splice |
+| `repeat` | main `!include`s the **same** parametrised leaf N times | jinja **recompilation** of an identical template |
+| `jinja` | one main file, `{% for %}` emitting N steps | single large render + single large parse |
+| `deep` | nested includes, depth N | include recursion (see caveat) |
+| `mix` | groups + jinja loop + distinct + repeated includes | realistic blend |
+
+## Reading the output
+
+```
+phase              min      median
+initial         0.1131      0.1285   <- pass 1: discover config files (no includes)
+loadtest        1.0724      1.0900   <- config fixpoint loop + full recursive include load
+build           0.1850      0.1976   <- TestSet: load_test_recursively tree build
+total           1.3886      1.4227
+counters  (last run):
+  templates :    1003 calls   0.5247s  (exclusive: jinja compile+render+tempfile)
+  yaml      :    1004 parses   1.4696s  (inclusive of nested includes)
+```
+
+- **min** is the headline (least noisy); median is a sanity check.
+- **initial / loadtest / build** map to the three pipeline stages in
+  `interpreter/process.py` and `interpreter/test_set.py`. The main file is
+  rendered+parsed across `initial` *and* `loadtest` (the loader does ~3 passes).
+- **templates** = number of `template_to_test()` calls and their *exclusive*
+  wall time (one file render each — pure jinja compile+render+tempfile I/O).
+  A high count with the same source file = recompilation, the `repeat` case.
+- **yaml** = number of `yaml_load()` parses. Its time is *inclusive* of nested
+  includes, so use the **count** for attribution, not the seconds.
+
+## Mapping to the optimisation axes
+
+| Axis (see DESIGN / discussion) | Watch | Best profile to prove it |
+|--------------------------------|-------|--------------------------|
+| 1 — cache compiled jinja templates | `templates` time drops, count unchanged | `repeat` |
+| 2 — drop the tempfile round-trip | `templates` time drops | `includes`, `repeat`, `mix` |
+| 3 — C YAML loader (libyaml) | `yaml` time / `loadtest` drops | `flat`, `jinja` |
+| 6 — O(n²) sequence splice | `build` drops | `includes`, `mix` |
+
+## How to compare before/after a change
+
+1. Run the matrix on the current code, keep the output.
+2. Apply one axis.
+3. Re-run the **same** profiles/sizes; compare `min` per phase and the counters.
+
+Change one axis at a time so the attribution is clean. Run on an idle machine
+(and note the disk: on a USB stick the tempfile round-trip of axis 2 weighs
+more).
+
+## Caveat: deep includes
+
+The loader is recursive and spends ~10 stack frames per include level, so
+`deep` hits Python's `RecursionError` around ~90 nested levels. The harness
+reports this cleanly instead of crashing. Real tests are *wide* (many steps /
+many includes), not deep, so `includes`/`repeat`/`jinja`/`mix` are the
+representative "very long" cases.
+
+## Notes
+
+- No execution is triggered — timing stops where `Batch` would mark the test
+  *loaded*.
+- The profiles contain no `<| |>`, so the external eval process is not started.
+  Pass `--with-eval` to `load_bench.py` for trees that evaluate at load time.
+- Numbers are machine- and disk-specific; only compare runs from the same host.

test/benchmark/gen_bench_test.py

@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+"""Generate synthetic ``.tum`` test trees to benchmark *load* time.
+
+The generated trees are deliberately cheap to *build* (only ``let`` leaves and
+``group`` containers — no subprocess, no runtime side effect) so the load
+benchmark measures the parse / template / tree-build pipeline and nothing else.
+
+Profiles, each targeting a specific cost in the loader:
+
+  flat      one main file, N inline ``let`` steps, no include, no jinja.
+            Baseline: YAML parse of a big document + linear object build.
+
+  includes  main ``!include``s N *distinct* sub-files (a few steps each).
+            Stresses the per-include template+YAML+tempfile round-trip and the
+            ``sequence`` splice in test_set.load_test_recursively.
+
+  repeat    main ``!include``s the *same* parametrised leaf file N times.
+            Stresses jinja *recompilation*: the compiled template is identical
+            every time, only the render params (idx) differ -> the case a
+            template cache collapses.
+
+  jinja     one main file whose ``{% for %}`` loop emits N steps.
+            Stresses a single large jinja render + a single large YAML parse.
+
+  deep      nested includes, depth N (main -> d0 -> d1 -> ...).
+            Stresses include recursion and per-level template+YAML.
+
+  mix       a realistic blend: groups, a jinja loop, distinct includes and a
+            repeated parametrised include.
+
+Usage:
+    gen_bench_test.py --profile repeat --size 1000 --out cases/repeat_1000
+    -> writes <out>/main.tum (+ includes, + param.yaml) and prints the path.
+"""
+import argparse
+import os
+import shutil
+
+
+def _let(indent, i, name=None):
+    name = name if name is not None else f"s{i}"
+    pad = " " * indent
+    return (
+        f"{pad}- let:\n"
+        f"{pad}    name: {name}\n"
+        f"{pad}    values:\n"
+        f"{pad}        - k{i}: {i}\n"
+    )
+
+
+def gen_flat(out, n):
+    body = "".join(_let(8, i) for i in range(n))
+    main = f"main:\n    name: bench flat {n}\n    steps:\n{body}"
+    _write(out, "main.tum", main)
+
+
+def gen_includes(out, n):
+    steps = "".join(f"        - !include inc_{i}.tum\n" for i in range(n))
+    main = f"main:\n    name: bench includes {n}\n    steps:\n{steps}"
+    _write(out, "main.tum", main)
+    for i in range(n):
+        # each include is a YAML *sequence* (list of steps)
+        seq = "".join(_let(0, i * 3 + j, name=f"inc{i}_{j}") for j in range(3))
+        _write(out, f"inc_{i}.tum", seq)
+
+
+def gen_repeat(out, n):
+    steps = "".join(
+        f"        - !include {{file: leaf.tum, idx: {i}}}\n" for i in range(n)
+    )
+    main = f"main:\n    name: bench repeat {n}\n    steps:\n{steps}"
+    _write(out, "main.tum", main)
+    leaf = (
+        "- let:\n"
+        "    name: leaf_{{ idx }}\n"
+        "    values:\n"
+        "        - leaf_{{ idx }}: {{ idx }}\n"
+    )
+    _write(out, "leaf.tum", leaf)
+
+
+def gen_jinja(out, n):
+    main = (
+        f"main:\n    name: bench jinja {n}\n    steps:\n"
+        "{% for i in range(" + str(n) + ") %}\n"
+        "        - let:\n"
+        "            name: j{{ i }}\n"
+        "            values:\n"
+        "                - k{{ i }}: {{ i }}\n"
+        "{% endfor %}\n"
+    )
+    _write(out, "main.tum", main)
+
+
+def gen_deep(out, n):
+    main = (
+        f"main:\n    name: bench deep {n}\n    steps:\n"
+        "        - let:\n            name: top\n            values:\n                - a: 0\n"
+        "        - !include d_0.tum\n"
+    )
+    _write(out, "main.tum", main)
+    for i in range(n):
+        seq = _let(0, i, name=f"d{i}")
+        if i < n - 1:
+            seq += f"- !include d_{i + 1}.tum\n"
+        _write(out, f"d_{i}.tum", seq)
+
+
+def gen_mix(out, n):
+    # n groups, each: 2 inline lets, one distinct include, one repeated include,
+    # plus a small jinja loop. Roughly ~6*n steps.
+    per = max(1, n)
+    parts = [f"main:\n    name: bench mix {n}\n    steps:\n"]
+    for g in range(per):
+        parts.append(
+            f"        - group:\n"
+            f"            name: grp{g}\n"
+            f"            steps:\n"
+        )
+        parts.append(_let(16, g * 2, name=f"g{g}_a"))
+        parts.append(_let(16, g * 2 + 1, name=f"g{g}_b"))
+        parts.append(f"                - !include inc_{g}.tum\n")
+        parts.append(f"                - !include {{file: leaf.tum, idx: {g}}}\n")
+        parts.append(
+            "{% for i in range(3) %}\n"
+            f"                - let:\n"
+            f"                    name: g{g}_j{{{{ i }}}}\n"
+            f"                    values:\n"
+            f"                        - g{g}_k{{{{ i }}}}: {{{{ i }}}}\n"
+            "{% endfor %}\n"
+        )
+    _write(out, "main.tum", "".join(parts))
+    for g in range(per):
+        _write(out, f"inc_{g}.tum", _let(0, g, name=f"mixinc{g}"))
+    _write(
+        out,
+        "leaf.tum",
+        "- let:\n    name: mixleaf_{{ idx }}\n    values:\n        - mixleaf_{{ idx }}: {{ idx }}\n",
+    )
+
+
+PROFILES = {
+    "flat": gen_flat,
+    "includes": gen_includes,
+    "repeat": gen_repeat,
+    "jinja": gen_jinja,
+    "deep": gen_deep,
+    "mix": gen_mix,
+}
+
+
+def _write(out, name, content):
+    with open(os.path.join(out, name), "w") as f:
+        f.write(content)
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--profile", required=True, choices=sorted(PROFILES))
+    ap.add_argument("--size", type=int, default=1000,
+                    help="profile-specific count (steps / includes / depth)")
+    ap.add_argument("--out", required=True, help="output directory (recreated)")
+    args = ap.parse_args()
+
+    out = os.path.abspath(args.out)
+    if os.path.isdir(out):
+        shutil.rmtree(out)
+    os.makedirs(out)
+
+    # minimal config file so the loader does not emit "no param file" noise
+    _write(out, "param.yaml", "bench_dummy: 1\n")
+
+    PROFILES[args.profile](out, args.size)
+    print(os.path.join(out, "main.tum"))
+
+
+if __name__ == "__main__":
+    main()

test/benchmark/load_bench.py

@@ -0,0 +1,200 @@
+#!/usr/bin/env python3
+"""Time the testium *load* pipeline on a given ``.tum`` tree.
+
+It drives the real loader code (``TestProcess._load_initial_params`` /
+``_load_test`` then ``TestSet(...)``) in-process, so the numbers track the
+production path and stay honest as the code evolves. Execution is never
+triggered — we stop exactly where ``Batch`` would report the test as *loaded*.
+
+Reported per run, over ``--repeat`` iterations (min is the headline, least
+noisy):
+
+    initial   first pass: discover config files (template+YAML, no includes)
+    loadtest  config-file fixpoint loop + full recursive include/template/YAML
+    build     TestSet construction: the load_test_recursively tree build
+    total     sum of the three
+
+Plus instrumentation counters (exact call counts, wall time) for the two
+hot leaves the optimisation axes target:
+
+    templates  jinja template_to_test() calls   (axis 1 compile cache, axis 2 tempfile)
+    yaml       yaml_load() parses               (axis 3 C loader)
+
+template time is exclusive (one file render); yaml time is wall-inclusive of
+nested includes, so lean on the *counts* for attribution.
+
+Must run inside the project venv (jinja2, pyyaml, telnetlib3, ...). The
+benchmark profiles contain no ``<| |>`` so the external eval process is not
+needed; pass --with-eval to start it for faithfulness on eval-heavy trees.
+
+Usage (see run.sh for the convenience wrapper):
+    test/tmp/.venv/bin/python3 test/benchmark/load_bench.py [--repeat 5] <main.tum>
+"""
+import argparse
+import os
+import statistics
+import sys
+from queue import Queue
+from time import perf_counter
+
+# --- bootstrap: src/testium for flat imports, src for `import testium` --------
+HERE = os.path.dirname(os.path.abspath(__file__))
+ROOT = os.path.abspath(os.path.join(HERE, "..", ".."))
+sys.path.insert(0, os.path.join(ROOT, "src"))
+sys.path.insert(0, os.path.join(ROOT, "src", "testium"))
+
+import api.testium as tm  # noqa: E402
+from interpreter.utils.test_init import env_init, apply_overrides  # noqa: E402
+from interpreter.utils.test_ctrl import TestSetController  # noqa: E402
+from interpreter.process import TestProcess  # noqa: E402
+from interpreter.test_set import TestSet  # noqa: E402
+from interpreter.utils.py_eval import eval_process_init  # noqa: E402
+from interpreter.utils.api_srv import api_request  # noqa: E402
+
+# --- instrumentation: count + time the two hot leaves -------------------------
+import interpreter.process as _proc  # noqa: E402
+import interpreter.utils.include as _inc  # noqa: E402
+import interpreter.utils.test_init as _ti  # noqa: E402
+import interpreter.utils.template as _tpl  # noqa: E402
+import interpreter.utils.yaml_load as _yl  # noqa: E402
+
+_C = {"tpl_n": 0, "tpl_t": 0.0, "yaml_n": 0, "yaml_t": 0.0}
+_orig_tpl = _tpl.template_to_test
+_orig_yaml = _yl.yaml_load
+
+
+def _wrap_tpl(*a, **k):
+    t = perf_counter()
+    try:
+        return _orig_tpl(*a, **k)
+    finally:
+        _C["tpl_t"] += perf_counter() - t
+        _C["tpl_n"] += 1
+
+
+def _wrap_yaml(*a, **k):
+    t = perf_counter()
+    try:
+        return _orig_yaml(*a, **k)
+    finally:
+        _C["yaml_t"] += perf_counter() - t
+        _C["yaml_n"] += 1
+
+
+# rebind in every module that did `from ... import template_to_test / yaml_load`
+for _m in (_proc, _inc):
+    _m.template_to_test = _wrap_tpl
+for _m in (_proc, _inc, _ti):
+    _m.yaml_load = _wrap_yaml
+
+
+def _reset_counters():
+    _C.update(tpl_n=0, tpl_t=0.0, yaml_n=0, yaml_t=0.0)
+
+
+def load_once(tp, fname, test_dir):
+    """One full load (no execution). Returns (initial, loadtest, build) seconds."""
+    t0 = perf_counter()
+    init_pf, gv = tp._load_initial_params(test_dir)
+    t1 = perf_counter()
+    test_dict, _pf = tp._load_test(init_pf, gv)
+    t2 = perf_counter()
+    TestSet(fname, test_dict, Queue())
+    t3 = perf_counter()
+    return (t1 - t0, t2 - t1, t3 - t2)
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("main_tum", help="path to the generated main.tum")
+    ap.add_argument("--repeat", type=int, default=5)
+    ap.add_argument("--with-eval", action="store_true",
+                    help="start the external eval process (needed only for <| |> at load)")
+    ap.add_argument("--quiet", action="store_true",
+                    help="silence the loader's INFO output during runs")
+    args = ap.parse_args()
+
+    fname = os.path.abspath(args.main_tum)
+    if not os.path.isfile(fname):
+        ap.error(f"not found: {fname}")
+    test_dir = os.path.dirname(fname)
+
+    env_init()
+    apply_overrides({}, {})
+
+    eval_proc = None
+    if args.with_eval:
+        eval_proc = eval_process_init(api_request, 10, test_dir)
+        eval_proc.start()
+        eval_proc.wait_ready(10)
+
+    if args.quiet:
+        # the loader prints a couple of INFO lines per config file; mute stdout
+        # around the measured section to avoid I/O skew.
+        devnull = open(os.devnull, "w")
+        real_stdout = sys.stdout
+
+    tp = TestProcess(fname, Queue(), TestSetController(),
+                     config_files=[], defines={}, gui_defaults={}, text_mode=True)
+
+    samples = []  # list of (initial, loadtest, build)
+    last_counters = None
+    try:
+        for r in range(args.repeat):
+            _reset_counters()
+            if args.quiet:
+                sys.stdout = devnull
+            try:
+                samples.append(load_once(tp, fname, test_dir))
+            except RecursionError:
+                if args.quiet:
+                    sys.stdout = real_stdout
+                print(f"file      : {fname}")
+                print("ERROR     : RecursionError during load — the include "
+                      "nesting is too deep for the recursive loader.\n"
+                      "            (each include level costs ~10 stack frames; "
+                      "raise sys.setrecursionlimit to probe further.)")
+                return 2
+            except Exception as e:  # noqa: BLE001 - report, don't crash the bench
+                if args.quiet:
+                    sys.stdout = real_stdout
+                print(f"file      : {fname}")
+                print(f"ERROR     : load failed: {type(e).__name__}: {e}")
+                return 2
+            finally:
+                if args.quiet:
+                    sys.stdout = real_stdout
+            last_counters = dict(_C)
+    finally:
+        if eval_proc is not None:
+            eval_proc.stop()
+            eval_proc.join()
+        if args.quiet:
+            devnull.close()
+
+    initial = [s[0] for s in samples]
+    loadtest = [s[1] for s in samples]
+    build = [s[2] for s in samples]
+    total = [sum(s) for s in samples]
+
+    def stat(xs):
+        return min(xs), statistics.median(xs)
+
+    print(f"file      : {fname}")
+    print(f"repeats   : {args.repeat}   (showing  min | median, seconds)")
+    print(f"{'phase':<10}{'min':>12}{'median':>12}")
+    for name, xs in (("initial", initial), ("loadtest", loadtest),
+                     ("build", build), ("total", total)):
+        mn, md = stat(xs)
+        print(f"{name:<10}{mn:>12.4f}{md:>12.4f}")
+    if last_counters:
+        print("counters  (last run):")
+        print(f"  templates : {last_counters['tpl_n']:>7d} calls   "
+              f"{last_counters['tpl_t']:>8.4f}s  (exclusive: jinja compile+render+tempfile)")
+        print(f"  yaml      : {last_counters['yaml_n']:>7d} parses  "
+              f"{last_counters['yaml_t']:>8.4f}s  (inclusive of nested includes)")
+
+
+if __name__ == "__main__":
+    sys.exit(main() or 0)

test/benchmark/run.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+# Load-time benchmark driver: generate synthetic .tum trees and time the
+# testium load pipeline on them, using the project venv.
+#
+# Usage:
+#   ./test/benchmark/run.sh                     # default matrix (all profiles)
+#   ./test/benchmark/run.sh <profile> <size>    # one profile at one size
+#   REPEAT=10 ./test/benchmark/run.sh repeat 2000
+#
+# Profiles: flat includes repeat jinja deep mix   (see gen_bench_test.py)
+#
+# Generated trees go under test/benchmark/cases/ (git-ignored). The numbers
+# are wall-clock; run on an otherwise idle machine and compare min values.
+set -e
+
+SCRIPT_DIR="$(realpath "$(dirname "$(readlink -f "$0")")")"
+PROJECT_DIR="$(realpath "$SCRIPT_DIR/../..")"
+VPY="$PROJECT_DIR/test/tmp/.venv/bin/python3"
+CASES="$SCRIPT_DIR/cases"
+REPEAT="${REPEAT:-5}"
+
+if [ ! -x "$VPY" ]; then
+    echo "ERROR: project venv not found at $VPY — run ./run.sh once to create it." >&2
+    exit 1
+fi
+
+bench() {
+    local profile="$1" size="$2"
+    local out="$CASES/${profile}_${size}"
+    local main
+    main="$("$VPY" "$SCRIPT_DIR/gen_bench_test.py" --profile "$profile" --size "$size" --out "$out")"
+    echo "===== profile=$profile size=$size ====="
+    "$VPY" "$SCRIPT_DIR/load_bench.py" --repeat "$REPEAT" --quiet "$main"
+    echo
+}
+
+if [ $# -eq 2 ]; then
+    bench "$1" "$2"
+    exit 0
+fi
+
+# Default matrix. 'deep' is kept small: the recursive loader hits Python's
+# recursion limit around ~90 nested include levels.
+bench flat     2000
+bench includes 1000
+bench repeat   1000
+bench jinja    2000
+bench deep     40
+bench mix      300

test/robustness/console serial/rsl_terminal_robustness.tum

@@ -1,69 +0,0 @@
-# Main
-################################################################################
-main:
-  name:                               Serial Terminal bug reproducer
-  version:                            0.1
-  steps:
-    - group:
-        name:                         Test preparation
-        steps:
-          - console:
-              name:             Open RSL Simulator Terminal
-              console_name:     RSL_simulator
-              steps:
-                  - open:
-                      protocol:   terminal
-                      terminal_path: $(rslsimulatorpath)
-                  - writeln: "pwd"
-                  - read_until: {expected: "$", timeout: 5}
-                  - writeln: "./RSverify $(rsTx)" # /dev/ttyMUE1
-                  - read_until: {expected: "RSL controller>", timeout: 5}
-                  - writeln: "setportconf 0 115200 none 8 1 1 255"
-                  - read_until: {expected: "RSL controller>", timeout: 5}
-                  - writeln: "send4ever 0 0"
-                  - read_until: {expected: "RSL controller>", timeout: 5}
-
-          - console:
-              name:                   Open the EUT console
-              console_name:           cons_target
-              doc:                    Initiates the console of the target in order
-                                      to be ready to capture its traces.
-              stop_on_failure:        True
-              steps:
-                - open:
-                    protocol:         serial
-                    serial_port:      $(rsRx) # /dev/ttyMUE2
-                    serial_baudrate:  115200
-
-    - loop:
-        name:                         Qualification loop
-        stop_on_failure:              False
-        steps:
-          - py_func:
-              name:                   Capture the RS serial output
-              file:                   $(test_directory)/terminal_bug_reproducer.py
-              func_name:              RetreiveData
-              param:
-                - cons_target
-
-          - sleep: {timeout: 1}
-
-# Cleanup sequence
-#-------------------------------------------------------------------------------
-    - group:
-        name:                         Cleanup
-        execute_on_stop:              True
-        steps:
-          - console:
-              name:                   Close the target console
-              console_name:           cons_target
-              execute_on_stop:        True
-              steps:
-                - close:
-
-          - console:
-              name:                   Close the RSL_simulator
-              console_name:           RSL_simulator
-              execute_on_stop:        True
-              steps:
-                - close:

test/robustness/console serial/terminal_bug_reproducer.py

@@ -1,26 +0,0 @@
-import api.testium as tm
-
-def RetreiveData(console_name):
-    print("--------------- retrieving data ---------------")
-    result = 0
-    cons   = tm.console(console_name)
-
-    if cons is None:
-        print("--------------- The console does not exist ---------------")
-    else:
-        try:
-            is_finished = False
-            while not is_finished:
-                status, d = cons.read_until('\n', timeout=0, return_data=True, mute=True)
-                if 0 == status:
-                    print("--------------- Data ---------------")
-                    print(d)
-                else:
-                    print("--------------- No data ---------------")
-                    print("Status: ", status)
-                    is_finished = True
-        except:
-            print("--------------- Error retrieving data ---------------")
-            result = -1
-
-    return result

test/robustness/console terminal/generate_char.sh

@@ -1,9 +0,0 @@
-chars='<=>| -,;:!/."()[]{}*\&#%+012345689abcdefghiklmnopqrstuvwxyzABCD'
-for j in {1..256} ;
-do
-    for i in {1..256} ; do
-        echo -n "${chars:RANDOM%${#chars}:1}"
-    done
-    echo
-    sleep 0.01
-done

test/robustness/console terminal/terminal_bug_reproducer.py

@@ -1,26 +0,0 @@
-import api.testium as tm
-
-def RetreiveData(console_name):
-    print("--------------- retrieving data ---------------")
-    result = 0
-    cons   = tm.console(console_name)
-
-    if cons is None:
-        print("--------------- The console does not exist ---------------")
-    else:
-        try:
-            is_finished = False
-            while not is_finished:
-                status, d = cons.read_until('\n', timeout=0, return_data=True, mute=True)
-                if 0 == status:
-                    print("--------------- Data ---------------")
-                    print(d)
-                else:
-                    print("--------------- No data ---------------")
-                    print("Status: ", status)
-                    is_finished = True
-        except:
-            print("--------------- Error retrieving data ---------------")
-            result = -1
-
-    return result

test/robustness/console terminal/terminal_robustness.tum

@@ -1,50 +0,0 @@
-# Main
-################################################################################
-main:
-  name:                               Terminal bug reproducer
-  version:                            0.1
-  steps:
-    - group:
-        name:                         Test preparation
-        steps:
-          - console:
-              name:                   Open the EUT console
-              console_name:           cons_target
-              doc:                    Initiates the console of the target in order
-                                      to be ready to capture its traces.
-              stop_on_failure:        True
-              steps:
-                - open:
-                    protocol: terminal
-
-    - loop:
-        name:                         Qualification loop
-        stop_on_failure:              False
-        steps:
-          - console:
-              name:                   write random data
-              console_name:           cons_target
-              steps:
-                  - writeln: bash $(test_directory)/generate_char.sh
-
-          - py_func:
-              name:                   Capture the terminal output
-              file:                   $(test_directory)/terminal_bug_reproducer.py
-              func_name:              RetreiveData
-              param:
-                - cons_target
-
-          - sleep: {timeout: 1}
-
-# Cleanup sequence
-#-------------------------------------------------------------------------------
-    - group:
-        name:                         Cleanup
-        execute_on_stop:              True
-        steps:
-          - console:
-              name:                   Close the target console
-              console_name:           cons_target
-              execute_on_stop:        True
-              steps:
-                - close:

test/validation/items/console/test.tum

@@ -84,7 +84,18 @@
         - read_until: {expected: HelloConsole, timeout: 1, mute: true}
 
 - console:
-    name: Console read_until muted
+    name: Console read_until float timeout
+    console_name: term
+    key: $(test)_PASS
+    steps:
+        - writeln: echo "HelloConsole"
+{% if os == "Windows" %}
+        - read_until: {expected: echo "HelloConsole", timeout: 0.2}
+{% endif %}
+        - read_until: {expected: HelloConsole, timeout: 0.2}
+
+- console:
+    name: Console read_until process result
     console_name: term
     key: $(test)_PASS
     steps:
@@ -94,6 +105,17 @@
 {% endif %}
         - read_until: {expected: endOfCmd, timeout: 1, process_result: "'Hello' in r'''$(result)''' and 'PASS' in r'''$(result)''' "}
 
+{% if os == "Linux" %}
+- console:
+    name: Console runs on host (not the Flatpak sandbox)
+    doc: Regression guard for the 0.2.1 Flatpak bug (term console spawned inside the sandbox instead of on the host). /.flatpak-info exists only inside the sandbox, so the host-only marker is emitted (and matched by read_until) ONLY when the shell really runs on the host. On a broken Flatpak the marker never appears, read_until times out and the item FAILS. The marker is built at runtime ($M) so it is never present in the command line itself. Passes on every other channel.
+    console_name: term
+    key: $(test)_PASS
+    steps:
+        - writeln: 'test -e /.flatpak-info && M=SANDBOX || M=HOST; echo "console_host_check_$M"'
+        - read_until: {expected: console_host_check_HOST, timeout: 5}
+{% endif %}
+
 - console:
     name: Console closure
     execute_on_stop: true

test/validation/items/let/test.tum

@@ -11,8 +11,8 @@
         - let:
             name:  Let it be
             values:
-                it: $(loop_param)
-                be: <| $(loop_param) == $(it) |>
+                - it: $(loop_param)
+                - be: <| $(loop_param) == $(it) |>
 
 - loop:
     name: Cycle iterating on list

test/validation/items/lua_func/test.tum

@@ -1,7 +1,7 @@
 - let:
     name: lua_func test constants,
     values:
-      lua_func test parameter: test parameter lua_func
+      - lua_func test parameter: test parameter lua_func
 
 - lua_func:
     name: fail lua_func

test/validation/items/py_func/test.tum

@@ -1,7 +1,7 @@
 - let:
     name: py_func test constants,
     values:
-      py_func test parameter: test parameter
+      - py_func test parameter: test parameter
 
 - py_func:
     name: pass py_func

test/validation/items/run/sub_fail.tum

@@ -1,7 +0,0 @@
-main:
-    name: run sub-test (always fail)
-    steps:
-        - check:
-            name: fail
-            values:
-                - false

test/validation/items/run/sub_pass.tum

@@ -1,7 +0,0 @@
-main:
-    name: run sub-test (always pass)
-    steps:
-        - check:
-            name: pass
-            values:
-                - true

test/validation/main.tum

@@ -31,7 +31,11 @@ main:
 
 {% for item in items %}
                 # item test
-                - let: {name: {{ item }} test constants, values: {test: {{ item }}, test_path: items/$(test)}}
+                - let: 
+                    name: {{ item }} test constants
+                    values: 
+                        - test: {{ item }}
+                        - test_path: items/$(test)
                 - group:
                     name: {{ item }} test
                     steps:

test/validation/run.sh

@@ -97,7 +97,10 @@ case "$MODE" in
             echo "Creating wheel venv at $WHEEL_VENV"
             python3 -m venv --system-site-packages "$WHEEL_VENV"
             "$WHEEL_VENV/bin/pip" install --quiet --upgrade pip
-            "$WHEEL_VENV/bin/pip" install --quiet "$WHEEL"
+            # Install with the [lsp] extra so the wheel channel is validated in
+            # its language-server-capable form (pulls pygls), matching how a
+            # user enables `testium lsp` from a wheel: pip install testium[lsp].
+            "$WHEEL_VENV/bin/pip" install --quiet "${WHEEL}[lsp]"
         fi
         CMD=("$WHEEL_VENV/bin/python" -m testium)
         ;;