@@ -40,6 +40,11 @@ test = [
     "pytest>=8.3.0,<9.0.0",
     "pytest-cov>=6.0.0",
     "httpx>=0.27.0,<0.29.0",
+    # Property-based tests for validate_input live in
+    # tests/pycharting/data/test_ingestion.py. Also supplied by rhiza's
+    # .rhiza/requirements/tests.txt, but declared here so `uv run pytest`
+    # works without `make install` having run first.
+    "hypothesis>=6.100.0",
 ]
 lint = [
     "ruff>=0.6.0",
 
@@ -0,0 +1,5 @@
+"""Performance benchmarks, run by ``make benchmark``.
+
+Excluded from ``make test`` (``--ignore=tests/benchmarks``) because benchmarks
+measure rather than assert, and are unreliable under ``pytest-xdist``.
+"""
@@ -0,0 +1,66 @@
+"""Benchmarks for the hot path: validation on ingest, and slicing on every viewport move.
+
+``DataManager.get_chunk`` is called once per pan/zoom in the browser, so its cost
+is what the user feels as lag. ``validate_input`` is paid once per ``plot()`` call
+but walks every series, so it dominates start-up on large inputs.
+
+Run with ``make benchmark``. These are excluded from ``make test``.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pandas as pd
+import pytest
+
+from pycharting.data.ingestion import DataManager, validate_input
+
+# Large enough to be dominated by the array work rather than call overhead,
+# small enough that the suite stays inside pytest.ini's 60s timeout.
+N = 250_000
+CHUNK = 5_000
+
+
+@pytest.fixture(scope="module")
+def ohlc() -> dict[str, np.ndarray]:
+    """A dense OHLC series of ``N`` bars with a datetime index and two overlays."""
+    rng = np.random.default_rng(seed=0)
+    close = np.cumsum(rng.standard_normal(N)) + 1_000.0
+    return {
+        "index": pd.date_range("2020-01-01", periods=N, freq="min"),
+        "open": close - 0.5,
+        "high": close + 1.0,
+        "low": close - 1.0,
+        "close": close,
+    }
+
+
+@pytest.fixture(scope="module")
+def manager(ohlc: dict[str, np.ndarray]) -> DataManager:
+    """A ``DataManager`` over the dense series, built once for the module."""
+    return DataManager(**ohlc)
+
+
+def test_validate_input_on_dense_ohlc(benchmark, ohlc: dict[str, np.ndarray]) -> None:
+    """Cost of normalizing a dense OHLC series — paid once per plot() call."""
+    result = benchmark(lambda: validate_input(**ohlc))
+    assert len(result["close"]) == N
+
+
+def test_get_chunk_viewport_slice(benchmark, manager: DataManager) -> None:
+    """Cost of one viewport slice — paid on every pan and zoom."""
+    mid = N // 2
+    result = benchmark(lambda: manager.get_chunk(mid, mid + CHUNK))
+    assert len(result["index"]) == CHUNK
+
+
+def test_get_chunk_full_range(benchmark, manager: DataManager) -> None:
+    """Cost of serializing the whole series, the worst case on first paint."""
+    result = benchmark(lambda: manager.get_chunk())
+    assert len(result["index"]) == N
+
+
+def test_data_manager_construction(benchmark, ohlc: dict[str, np.ndarray]) -> None:
+    """End-to-end ingest cost: validation plus array retention."""
+    result = benchmark(lambda: DataManager(**ohlc))
+    assert result.length == N
@@ -6,6 +6,9 @@
 import numpy as np
 import pandas as pd
 import pytest
+from hypothesis import given, settings
+from hypothesis import strategies as st
+from hypothesis.extra.numpy import arrays
 
 from pycharting.data.ingestion import DataManager, DataValidationError, validate_input
 
@@ -779,3 +782,80 @@ def test_get_chunk_data_types(self):
 
         # Individual values should be Python numbers
         assert isinstance(chunk["open"][0], (int, float))
+
+
+# ---------------------------------------------------------------------------
+# Property-based tests — run by `make hypothesis-test` (-m "hypothesis or property")
+# and, being under tests/pycharting/, also by the regular `make test` run.
+#
+# validate_input's contract is stated as invariants rather than examples: every
+# output is a length-n ndarray, length disagreement always raises, and the
+# single-vs-multi series mode is decided purely by how many of OHLC are present.
+# Those are properties, so they are tested as properties.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.property
+@given(
+    values=arrays(
+        dtype=np.float64,
+        shape=st.integers(min_value=1, max_value=200),
+        elements=st.floats(min_value=-1e6, max_value=1e6, allow_nan=False),
+    ),
+)
+@settings(max_examples=50, deadline=None)
+def test_close_only_input_always_normalizes_to_line_mode(values):
+    """Any single series is mapped to `close`, leaving open/high/low unset."""
+    result = validate_input(np.arange(len(values)), close=values)
+
+    assert isinstance(result["close"], np.ndarray)
+    assert len(result["close"]) == len(values)
+    assert result["open"] is None
+    assert result["high"] is None
+    assert result["low"] is None
+
+
+@pytest.mark.property
+@given(
+    n=st.integers(min_value=1, max_value=200),
+    delta=st.integers(min_value=1, max_value=20),
+)
+@settings(max_examples=50, deadline=None)
+def test_length_mismatch_always_raises(n, delta):
+    """A series whose length differs from the index is always rejected."""
+    index = np.arange(n)
+    close = np.zeros(n + delta)
+
+    with pytest.raises(DataValidationError, match="does not match index length"):
+        validate_input(index, close=close)
+
+
+@pytest.mark.property
+@given(
+    values=arrays(
+        dtype=np.float64,
+        shape=st.integers(min_value=1, max_value=100),
+        elements=st.floats(min_value=1.0, max_value=1e5, allow_nan=False),
+    ),
+)
+@settings(max_examples=50, deadline=None)
+def test_ohlc_high_is_never_below_low(values):
+    """With open and close supplied, the auto-filled high never falls below low."""
+    result = validate_input(np.arange(len(values)), open=values, close=values + 1.0)
+
+    assert np.all(result["high"] >= result["low"])
+
+
+@pytest.mark.property
+@given(
+    values=st.lists(st.floats(min_value=-1e6, max_value=1e6, allow_nan=False), min_size=1, max_size=100),
+)
+@settings(max_examples=50, deadline=None)
+def test_list_series_and_ndarray_series_agree(values):
+    """Passing a list and the equivalent ndarray produce the same normalized output."""
+    index = np.arange(len(values))
+
+    from_list = validate_input(index, close=values)
+    from_array = validate_input(index, close=np.array(values, dtype=np.float64))
+
+    assert np.array_equal(from_list["close"], from_array["close"])
@@ -0,0 +1,6 @@
+"""Stress and load tests, run by ``make stress`` (selected via ``-m stress``).
+
+Excluded from ``make test`` (``--ignore=tests/stress``) because these push
+volume and concurrency rather than assert a single behaviour, and are slower
+than the unit suite.
+"""
@@ -0,0 +1,108 @@
+"""Stress tests for the session registry and the slicing path under concurrency.
+
+``_data_managers`` is a plain module-level dict shared by every request handler,
+and the chart frontend issues overlapping ``/api/data`` requests while the user
+pans. These tests push that path harder than the unit suite does: many sessions
+at once, churn of create/delete, and concurrent reads of the same manager.
+
+Run with ``make stress``. Excluded from ``make test``.
+"""
+
+from __future__ import annotations
+
+from concurrent.futures import ThreadPoolExecutor
+
+import numpy as np
+import pytest
+from fastapi.testclient import TestClient
+
+from pycharting.api.routes import _data_managers
+from pycharting.core.server import create_app
+from pycharting.data.ingestion import DataManager
+
+pytestmark = pytest.mark.stress
+
+BARS = 20_000
+SESSIONS = 100
+WORKERS = 16
+
+
+@pytest.fixture
+def client() -> TestClient:
+    """A test client over a fresh app, with the session registry cleared around it."""
+    _data_managers.clear()
+    try:
+        yield TestClient(create_app())
+    finally:
+        _data_managers.clear()
+
+
+def _series(n: int = BARS) -> dict[str, np.ndarray]:
+    """A deterministic OHLC payload of ``n`` bars."""
+    rng = np.random.default_rng(seed=1)
+    close = np.cumsum(rng.standard_normal(n)) + 500.0
+    return {
+        "index": np.arange(n),
+        "open": close - 0.5,
+        "high": close + 1.0,
+        "low": close - 1.0,
+        "close": close,
+    }
+
+
+def test_many_concurrent_sessions_stay_isolated(client: TestClient) -> None:
+    """Registering many sessions at once keeps each one's data distinct."""
+    payload = _series(1_000)
+
+    def register(i: int) -> str:
+        # Shift the whole bar, not just close — validate_input enforces
+        # high >= max(open, close), so offsetting one series in isolation
+        # is rejected.
+        session = f"stress-{i}"
+        shifted = {k: (v + i if k != "index" else v) for k, v in payload.items()}
+        _data_managers[session] = DataManager(**shifted)
+        return session
+
+    with ThreadPoolExecutor(max_workers=WORKERS) as pool:
+        sessions = list(pool.map(register, range(SESSIONS)))
+
+    assert len(_data_managers) == SESSIONS
+    # Each session kept its own offset rather than aliasing a shared array.
+    for i, session in enumerate(sessions):
+        assert _data_managers[session].close[0] == pytest.approx(payload["close"][0] + i)
+
+
+def test_concurrent_chunk_reads_are_consistent(client: TestClient) -> None:
+    """Overlapping viewport reads of one session all return the same bytes."""
+    _data_managers["shared"] = DataManager(**_series())
+    manager = _data_managers["shared"]
+    expected = manager.get_chunk(0, 500)["close"]
+
+    def read(_: int) -> list[float]:
+        return manager.get_chunk(0, 500)["close"]
+
+    with ThreadPoolExecutor(max_workers=WORKERS) as pool:
+        results = list(pool.map(read, range(200)))
+
+    assert all(r == expected for r in results)
+
+
+def test_session_churn_does_not_leak(client: TestClient) -> None:
+    """Repeated create/delete cycles leave the registry empty."""
+    for i in range(SESSIONS):
+        session = f"churn-{i}"
+        _data_managers[session] = DataManager(**_series(500))
+        response = client.delete(f"/api/sessions/{session}")
+        assert response.status_code == 200
+
+    assert _data_managers == {}
+
+
+def test_full_range_slice_of_a_large_series(client: TestClient) -> None:
+    """A whole-series slice serializes every bar without truncating."""
+    _data_managers["big"] = DataManager(**_series())
+
+    chunk = _data_managers["big"].get_chunk()
+
+    assert len(chunk["index"]) == BARS
+    assert len(chunk["close"]) == BARS