chore: add opt-in large-tier benchmarks

benchmarks/README.md

@@ -24,6 +24,17 @@ uv run --with "pytest-benchmem[plot]" pytest benchmarks/ \
 (`-o addopts="" -p no:xdist` because the project's xdist config auto-disables
 benchmarks.)
 
+## Large tier
+
+Production-sized cases (`test_large_*`, several seconds and hundreds of MiB
+peak each) are skipped by default — in CI too. Run them locally before
+dependency bumps or performance work:
+
+```bash
+uv run --with "pytest-benchmem[plot]" pytest benchmarks/ --large -k large \
+    -o addopts="" -p no:xdist --benchmark-only --benchmark-memory
+```
+
 ## Baselines
 
 Save a named baseline (stored in `.benchmarks/`, gitignored — machine-specific):

benchmarks/conftest.py

@@ -0,0 +1,25 @@
+import pytest
+
+
+def pytest_addoption(parser):
+    parser.addoption(
+        "--large",
+        action="store_true",
+        default=False,
+        help="run large-scale benchmarks (production-sized data, minutes)",
+    )
+
+
+def pytest_configure(config):
+    config.addinivalue_line(
+        "markers", "large: large-scale benchmark, runs only with --large"
+    )
+
+
+def pytest_collection_modifyitems(config, items):
+    if config.getoption("--large"):
+        return
+    skip = pytest.mark.skip(reason="large benchmark: pass --large to run")
+    for item in items:
+        if "large" in item.keywords:
+            item.add_marker(skip)

benchmarks/test_bench_aggregate.py

@@ -17,6 +17,10 @@
   integration surprises (tsam version bumps, config plumbing), the
   representation-by-width interaction, and the multi-``cluster_dim``
   MultiIndex path.
+- ``test_large_*`` — production-sized cases (hundreds of MiB peak) that
+  guard the memory-amplification regime. Opt-in via ``--large`` (skipped
+  otherwise, including in CI) — run locally before dependency bumps or
+  performance work.
 - ``test_user_*`` — post-aggregation operations users run on results:
   reusing a stored clustering on new data (``clustering.apply``) and
   expanding cluster-level data back to the full time axis
@@ -226,3 +230,20 @@ def test_user_disaggregate(benchmark, wide_result):
         args=(result.cluster_representatives,),
         **CONFIG_OPTS,
     )
+
+
+# --- large tier (production-sized, opt-in via --large) -------------------------
+
+LARGE_OPTS = dict(rounds=2, iterations=1, warmup_rounds=0)
+
+
+@pytest.mark.large
+def test_large_wide(benchmark):
+    da = make_data(730, n_cols=256, n_slices=1)
+    benchmark.pedantic(run_aggregate, args=(da, FULL_CONFIG), **LARGE_OPTS)
+
+
+@pytest.mark.large
+def test_large_scenarios(benchmark):
+    da = make_data(365, n_cols=64, n_slices=8)
+    benchmark.pedantic(run_aggregate, args=(da, FULL_CONFIG), **LARGE_OPTS)