fix: remove background model preload (degraded perf on cpu-basic)

The daemon-thread preload competed for HF cpu-basic's 2 vCPUs and held a
load lock, so page reloads blocked behind it and felt very slow — the
opposite of the goal. Drop it entirely and reduce model_helpers to the
plain @st .cache_resource loader, which already guarantees a single
process-wide load reused across sessions/reruns with nothing that can
block a reload. Keeps the real win: the parser and custom-label demos no
longer reload the model on every interaction.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -1,17 +1,10 @@
 import streamlit as st
 
-from model_helpers import start_preload
-
 st.set_page_config(
     page_title="LatinCy Dashboard | Home",
     page_icon="🏠",
 )
 
-# Warm the default lg model in a background thread while the visitor reads this
-# page, so the first demo they open finds it already loaded. Non-blocking and
-# idempotent — safe to call on every render.
-start_preload()
-
 st.write("# LatinCy Dashboard")
 
 st.sidebar.success("Select a demo above.")

model_helpers.py

@@ -1,63 +1,30 @@
-"""Shared spaCy model loading + background preload for the dashboard.
+"""Shared spaCy model loading for the dashboard.
 
-Read-only demos import ``load_model()`` from here so they all share ONE model
+Read-only demos import ``load_model()`` so they all share ONE cached model
 instance per name across the whole process, instead of each page caching its
 own copy (which multiplied a ~500 MB model across pages in RAM).
-``start_preload()`` warms the default model in a background thread from the
-home page, so navigating into a demo finds the model already resident.
+``@st.cache_resource`` already guarantees the model loads once per process and
+is reused across sessions/reruns — no manual threading needed.
 
 Demos that MUTATE their pipeline (e.g. the custom-label demo, which adds a
-component with ``add_pipe``) must NOT use this shared instance — they should
-cache their own copy — because mutating the shared model would leak the change
-into every other demo.
+component with ``add_pipe``) must cache their OWN instance instead — mutating
+this shared one would leak the change into every other demo.
 """
 
-import threading
-
 import spacy
 import streamlit as st
 
 DEFAULT_MODEL = "la_core_web_lg"
 
-# Single source of truth for loaded models, shared between the background
-# preloader and the synchronous loader. Guarded by _lock so a preload and a
-# concurrent load_model() can never start two loads of the same model.
-_models = {}
-_lock = threading.Lock()
-
-
-def _load(model_name):
-    """Load a model with no Streamlit (st.*) calls — safe off the main thread."""
-    nlp = spacy.load(model_name)
-    if "trf_vectors" in nlp.pipe_names:  # heavy, unused in these demos
-        nlp.disable_pipe("trf_vectors")
-    return nlp
-
-
-def _get(model_name):
-    with _lock:
-        if model_name not in _models:
-            _models[model_name] = _load(model_name)
-        return _models[model_name]
-
-
-def start_preload(model_name=DEFAULT_MODEL):
-    """Kick off a non-blocking background load of the default model.
-
-    Called from the home page so the model warms while the user reads the demo
-    list. Cheap to call repeatedly: once loaded, the worker returns at once.
-    """
-    if model_name in _models:
-        return
-    threading.Thread(target=_get, args=(model_name,), daemon=True).start()
-
 
 @st.cache_resource(show_spinner="Loading LatinCy model…")
 def load_model(model_name=DEFAULT_MODEL):
     """Return the shared, cached spaCy model for ``model_name``.
 
-    Reuses the background-preloaded instance when ready; otherwise loads it once
-    synchronously. Read-only use only — do not ``add_pipe``/``disable`` on the
-    result, since every demo shares this instance.
+    Read-only use only — do not ``add_pipe``/``disable`` on the result, since
+    every read-only demo shares this one instance.
     """
-    return _get(model_name)
+    nlp = spacy.load(model_name)
+    if "trf_vectors" in nlp.pipe_names:  # heavy, unused in these demos
+        nlp.disable_pipe("trf_vectors")
+    return nlp