fix(hooks): stop the pre-push gauntlet hanging on model-download tests

The hook ran the full suite, which includes tests that drive the real docling
pipeline and reranker. On a cold cache those download HF models (figure
classifier, CrossEncoder), stalling `git push` for minutes with no visible
output (the hook redirects pytest to a logfile). Mark those tests slow and run
the fast subset in the hook (-m "not slow and not integration"); slow and
integration still run in full in CI.

Author: NorthernLightx

.githooks/pre-push

@@ -39,9 +39,9 @@ PYTEST="$VENV_DIR/pytest"
 
 printf "pre-push gauntlet:\n"
 
-# Match the gates in .github/workflows/ci.yml exactly — running a narrower
-# subset means CI fails at the runner level on things the hook should have
-# caught (which is exactly what we're trying to prevent).
+# Lint, format, and types match .github/workflows/ci.yml exactly. The pytest
+# step runs only the fast subset; slow / integration tests pull model downloads
+# or docker-compose services and run in CI, not in this quick local gate.
 
 # 1. Ruff lint
 if "$RUFF" check . >/dev/null; then
@@ -67,8 +67,8 @@ else
     fail "mypy strict"
 fi
 
-# 4. Full pytest
-if "$PYTEST" tests/ -q --no-header >/tmp/pre-push-pytest.log 2>&1; then
+# 4. Fast pytest subset (slow + integration run in CI, not here)
+if "$PYTEST" tests/ -m "not slow and not integration" -q --no-header >/tmp/pre-push-pytest.log 2>&1; then
     pass "pytest ($(grep -oE '[0-9]+ passed' /tmp/pre-push-pytest.log | head -1))"
 else
     cat /tmp/pre-push-pytest.log

tests/unit/test_ingestion_pipeline.py

@@ -6,6 +6,7 @@
 from typing import Any
 
 import fitz
+import pytest
 
 from src.ingestion.pipeline import IngestedPaper, ingest_paper
 from src.llm.protocol import ChatResponse, Message
@@ -14,6 +15,10 @@
 from src.types import Paper
 from tests.fakes import FakeEmbedder
 
+# Real docling conversion downloads/loads a HF figure-classifier model — too
+# heavy for the fast pre-push gauntlet. Runs in CI, skipped by the local hook.
+pytestmark = pytest.mark.slow
+
 
 class _StubLLM:
     def __init__(self, text: str = "blurb") -> None:

tests/unit/test_main_wiring.py

@@ -148,6 +148,7 @@ async def test_wire_populates_pipeline_retriever_from_qdrant() -> None:
 
 
 @pytest.mark.asyncio
+@pytest.mark.slow  # real wiring reranks → CrossEncoder model download; CI only
 async def test_wired_retriever_can_serve_a_query() -> None:
     """Sanity check beyond construction: the wired retriever actually returns
     chunks (so set_retriever didn't silently register a broken instance)."""
@@ -347,6 +348,7 @@ async def test_multimodal_on_without_visual_leg_degrades_to_text() -> None:
 
 
 @pytest.mark.asyncio
+@pytest.mark.slow  # real wiring reranks → CrossEncoder model download; CI only
 async def test_multimodal_routing_actually_dispatches_figure_to_hybrid(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None: