Slim the slow checks: CPU-torch CI installs, a one-file GPU suite, no provenance-only sidecar fields

.github/workflows/ci.yml

@@ -28,26 +28,49 @@ jobs:
       - uses: actions/setup-python@v7
         with:
           python-version: "3.12"
-          cache: pip
-          cache-dependency-path: pyproject.toml
-      - run: python -m pip install --upgrade pip
-      # Install dev extras so pyright can resolve runtime/test dependencies
-      # while it analyzes the package, tests, examples, and maintained scripts.
-      - run: pip install -e ".[dev]"
+      - uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e # v6.8.0
+        with:
+          enable-cache: true
+          cache-dependency-glob: pyproject.toml
+      # CPU torch: the default Linux wheel drags in ~2.5 GB of CUDA libraries
+      # that neither pyright nor the CPU suite touches.
+      - run: uv pip install --system torch --index-url https://download.pytorch.org/whl/cpu
+      - run: uv pip install --system -e ".[dev]"
       - run: pyright
 
-  test:
+  # The CPU suite in four parallel shards (a runner has 4 vCPUs; the suite is
+  # ~330 CPU-seconds locally, a third of it the route tests).  The last shard
+  # takes every directory the others don't name, so a new one is never skipped.
+  test-shard:
     runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        shard:
+          - "tests/server"
+          - "tests/lens tests/units tests/sae tests/cli"
+          - "tests/fit tests/session tests/model tests/browser tests/loom"
+          - "tests --ignore=tests/server --ignore=tests/lens --ignore=tests/units --ignore=tests/sae --ignore=tests/cli --ignore=tests/fit --ignore=tests/session --ignore=tests/model --ignore=tests/browser --ignore=tests/loom"
     steps:
       - uses: actions/checkout@v7
       - uses: actions/setup-python@v7
         with:
           python-version: "3.13"
-          cache: pip
-          cache-dependency-path: pyproject.toml
-      - run: python -m pip install --upgrade pip
-      - run: pip install -e ".[dev]"
-      - run: pytest -q -m "not gpu" -n auto
+      - uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e # v6.8.0
+        with:
+          enable-cache: true
+          cache-dependency-glob: pyproject.toml
+      - run: uv pip install --system torch --index-url https://download.pytorch.org/whl/cpu
+      - run: uv pip install --system -e ".[dev]"
+      - run: pytest -q -m "not gpu" -n auto ${{ matrix.shard }}
+
+  # The one check branch protection names.
+  test:
+    if: always()
+    needs: test-shard
+    runs-on: ubuntu-latest
+    steps:
+      - run: test "${{ needs.test-shard.result }}" = success
 
   webui:
     runs-on: ubuntu-latest

AGENTS.md

@@ -209,11 +209,12 @@ get wrong raises `RequestError` (400, text verbatim) or a subclass
 failures after parsing are `SteeringCompositionError` (422). A bare
 `ValueError` is a broken invariant: the server logs it and answers a scrubbed
 500. The wire `code` is the exception class name. The CLI maps 404 → exit 1,
-other 4xx → 2, 5xx → 1. `tests/invariants/test_user_message.py` enforces the
-family.
+other 4xx → 2, 5xx → 1.
 
-**Persisted formats are exact.** Readers accept only the current version and
-never migrate (a derived cache recomputes instead). These froze at 6.0.0:
+**Persisted formats are versioned.** Readers accept only the current version and
+never migrate (a derived cache recomputes instead); the fitted sidecar reader
+ignores keys the contract has retired, so dropping provenance never strands a
+fit. These froze at 6.0.0:
 `manifold.json` v10, `template.json` v2, `.drowse` v1, loom `tree_format` 2
 with token sidecar 3 (token rows carry measurement envelope v2; a node's
 `finish_reason` is one of `stop`, `length`, `stop_sequence`, `cancelled`),
@@ -309,10 +310,11 @@ fit loop, `lens/fit.py`, is the reference pattern).
 tiny generated checkpoint (`tiny_session`); route tests that need no behavior
 use the autospec `tests.support.server.mock_session`. GPU tests (`gpu` marker)
 load `HuggingFaceTB/SmolLM2-360M-Instruct` unless `DROWSE_TEST_MODEL` names
-another; `gpu_gemma` needs gated Gemma access. Every bug fix lands with a
-regression test. A test never touches the real home (an autouse fixture sets a
-scratch `DROWSE_HOME`). CI runs ruff, pyright, the CPU suite on Python 3.13
-(`-n auto`) and the webui build with a dist-drift check. Details: `tests/AGENTS.md`.
+another; the GPU suite is one end-to-end smoke plus the throughput gates. Add a
+regression test when a bug could quietly come back, not as a ritual. A test
+never touches the real home (an autouse fixture sets a scratch `DROWSE_HOME`). CI runs ruff, pyright, the CPU suite on Python 3.13
+(four parallel shards, `-n auto`, CPU torch via uv) and the webui build with a
+dist-drift check. Details: `tests/AGENTS.md`.
 
 ## State and environment
 
@@ -333,4 +335,4 @@ downloads, and Python saves only to an explicit `LoomTree.save(path)`.
 | `DROWSE_STRICT_MODEL` | OpenAI/Ollama 404 on a model-name mismatch |
 | `DROWSE_MAX_REQUEST_BYTES` | request body cap (default 64 MiB) |
 | `DROWSE_MANIFOLD_CAPTURE_CACHE_GB` | fit capture cache bound (default 8) |
-| `DROWSE_TEST_MODEL`, `DROWSE_TEST_SAELENS`, `DROWSE_TEST_LIVE_TOKENIZERS` | test knobs (`tests/AGENTS.md`) |
+| `DROWSE_TEST_MODEL`, `DROWSE_TEST_LIVE_TOKENIZERS` | test knobs (`tests/AGENTS.md`) |

drowse/fit/AGENTS.md

@@ -375,8 +375,7 @@ corpora `tests/session/test_generate_responses.py`, session bake
 `tests/session/test_session_bake.py`, neutral cache
 `tests/store/test_neutral_cache.py`, neutral unit statistics
 `tests/units/test_unit_stats.py`, DLS golden
-`tests/browser/test_browser_parity.py`, GPU `tests/gpu/test_smoke.py` and
-`tests/gpu/test_session.py`.
+`tests/browser/test_browser_parity.py`, GPU `tests/gpu/test_smoke.py`.
 
 ## Deliberately absent
 

drowse/fit/pipeline.py

@@ -798,16 +798,9 @@ def accepts(self, sidecar: "ManifoldSidecar") -> bool:
         return bool(sidecar.sae_full_coverage)
 
     def base_metadata(
-        self, *, method: str, fitted_layers: list[int], subspace_metric: str,
+        self, *, method: str, fitted_layers: list[int],
     ) -> dict[str, Any]:
-        """The sidecar keys every fit writes.
-
-        ``subspace_metric`` is a basis *label*: the monopolar fold keeps the
-        raw δ̂ basis (``concept − ν`` cancels common-mode by differencing, like
-        DiM) and reports ``euclidean``; every other fit selects a whitened
-        basis.  The share is always Mahalanobis — the whitener gate guarantees
-        it.
-        """
+        """The sidecar keys every fit writes."""
         ident = self.identity
         metadata: dict[str, Any] = {
             "method": method,
@@ -820,9 +813,6 @@ def base_metadata(
             "baseline_prompts_sha256": ident.baseline_prompts_sha256,
             "fitted_layers": fitted_layers,
             "fit_policy_version": MANIFOLD_FIT_POLICY_VERSION,
-            # Provenance only (nothing branches on these at load).
-            "share_metric": "mahalanobis",
-            "subspace_metric": subspace_metric,
         }
         backend = self.sae_backend
         if backend is not None:
@@ -1342,7 +1332,6 @@ def _fit_monopolar(
                 else "manifold_monopolar"
             ),
             fitted_layers=list(fit_layers),
-            subspace_metric="euclidean",
         )
         metadata.update({
             "fit_mode": mf.fit_mode,
@@ -1448,7 +1437,6 @@ def _fit_nodes(
         metadata = inputs.base_metadata(
             method=geometry.method,
             fitted_layers=sorted(layer_subs),
-            subspace_metric="mahalanobis",
         )
         # Diagnostic layer profile: the whitened between-node spread
         # ``tr(G_L)`` — surfaced by `manifold show`, consumed by nothing.

drowse/geometry/fold.py

@@ -130,7 +130,6 @@ def fold_directions_to_subspace(
         node_kinds=[None],
         mahalanobis_share=mahalanobis_share,
     )
-    manifold.metadata["share_metric"] = "mahalanobis"
     manifold.validate_runtime_geometry()
     return manifold
 

drowse/lens/AGENTS.md

@@ -733,5 +733,3 @@ the per-layer softmax probability `p_l(v)`.
   (`LensHarness`, `lens_instrument(**kw)`, `lens_kwargs(**kw)`,
   `FakeCapture`), which builds a real `LensInstrument` with session-shaped
   dependencies. Generation-boundary tests use `tiny_session`.
-  `tests/gpu/test_jlens_gpu.py` micro-fits a real model (SmolLM2-360M or
-  `DROWSE_TEST_MODEL`).

drowse/model/AGENTS.md

@@ -446,5 +446,6 @@ and stitcher against the Jinja fakes in `tests/support/chat_templates.py`
 `test_role_template_drift.py` checks `SEAT_HEADERS` against real tokenizers in
 the local HF cache; `DROWSE_TEST_LIVE_TOKENIZERS=1` permits downloads (run it
 before a release). `tests/support/tiny_models.py` supplies genuine 4-layer
-models with real family templates for CPU end-to-end tests; real-weight checks
-are `tests/gpu/test_model_numerics_gpu.py` (`DROWSE_TEST_MODEL`).
+models with real family templates for CPU end-to-end tests;
+`tests/model/test_model_numerics.py` carries `mps` params for real-device
+numerics.

drowse/readout/AGENTS.md

@@ -284,8 +284,8 @@ rows at its own rank, one batched foot solve (`invert_stacked`) per rank
 (one for a rank-uniform fit), and the share-weighted means as a left fold in
 layer order. Every kernel computes a layer as the per-layer read did, so a
 reading does not depend on how many layers ran together: bit-identical to the
-per-layer read on MPS (`tests/gpu/test_curved_stack_gpu.py`; on CPU
-`tests/readout/test_monitor_curved_stack.py` pins it to tolerance). A
+per-layer read (`tests/readout/test_monitor_curved_stack.py` pins it to
+tolerance on CPU). A
 reduction batched across layers moves values by an ulp, and an off-surface
 state's foot to another basin, so keep reductions per layer. A read that saw only some of
 a probe's layers uses a cached `_CurvedStack.subset`. Curved feet warm-start

drowse/sae/AGENTS.md

@@ -346,5 +346,4 @@ instrument). `tests/support/sae.py` provides `MockSaeBackend`,
 section and recorded registry effects) and `make_resident(runtime, backend,
 layer=...)`. Session-level tests load a mock through `session.sae.load` with
 `drowse.sae.runtime.resolve_sae_backend` monkeypatched. Tests that need SAELens
-`importorskip` it. The real end-to-end run, `tests/gpu/test_sae_lens_smoke.py`,
-is opt-in via `DROWSE_TEST_SAELENS=1` and is marked `gpu_gemma`.
+`importorskip` it.

drowse/session/AGENTS.md

@@ -393,6 +393,6 @@ the root `AGENTS.md` "Errors" rule; `user_message()` defaults to
   `tests.support.engine.engine_over`, `tests.support.extraction.make_extraction`.
   Route tests use `tests.support.server.mock_session`, an autospec that fails on
   removed or renamed members.
-- GPU: `tests/gpu/test_session.py`, and `tests/gpu/test_smoke.py` (owns the
-  throughput regression), on `HuggingFaceTB/SmolLM2-360M-Instruct` unless
-  `DROWSE_TEST_MODEL` names another model.
+- GPU: `tests/gpu/test_smoke.py` (one end-to-end smoke and the throughput
+  gates), on `HuggingFaceTB/SmolLM2-360M-Instruct` unless `DROWSE_TEST_MODEL`
+  names another model.

drowse/session/extraction.py

@@ -393,7 +393,6 @@ def _fit_vector(
             metadata={
                 "method": "manifold_pca",
                 "name": ret_name,
-                "share_metric": manifold.metadata.get("share_metric"),
             },
         )
         self._events.emit(ManifoldExtracted(

drowse/store/AGENTS.md

@@ -88,7 +88,7 @@ provenance only, and an unknown key raises. Provenance is all-or-nothing against
 | `method` | `fit_mode` | required block |
 | --- | --- | --- |
 | `manifold_pca`, `manifold_sae` | authored | |
-| `manifold_monopolar[_sae]` | pca | 1-node ray, `subspace_metric = euclidean` |
+| `manifold_monopolar[_sae]` | pca | 1-node ray (`concept − ν`) |
 | `manifold_discover_{pca,spectral,auto}` | its own | auto: `resolved_fit_mode`, `topology_winner`, `topology_candidates` |
 | `manifold_discover_sae` | any discover | |
 | `manifold_procrustes_transfer` | any | `source_model_id`, `source_model_fingerprint` |

drowse/store/manifold_authoring.py

@@ -453,9 +453,6 @@ def build(staging: Path) -> None:
                 "method": method, "fit_mode": "baked", "nodes_sha256": nodes_sha,
                 "model_fingerprint": item.model_fingerprint,
             }
-            share_metric = item.manifold.metadata.get("share_metric")
-            if share_metric:
-                meta["share_metric"] = share_metric
             if item.components:
                 meta["components"] = item.components
             if method == "merge":

drowse/store/manifold_lifecycle.py

@@ -457,10 +457,6 @@ def transfer_manifold(
             metadata["source_model_id"] = from_model
             metadata["source_model_fingerprint"] = source_model_fingerprint
             metadata["model_fingerprint"] = target_model_fingerprint
-            # The share was re-baked in the target metric; ``subspace_metric``
-            # keeps the source value (the basis was selected there and only
-            # rotated here).
-            metadata["share_metric"] = "mahalanobis"
             if transfer_quality_estimate is not None:
                 metadata["transfer_quality_estimate"] = float(transfer_quality_estimate)
             sidecar_path = out_path.with_suffix(".json")
@@ -502,8 +498,6 @@ def manifold_fit_summary(sidecar: ManifoldSidecar, stem: str) -> dict[str, Any]:
         ("hyperparams", sidecar.hyperparams),
         ("diagnostics", sidecar.diagnostics),
         ("node_spread", sidecar.node_spread_per_layer),
-        ("share_metric", sidecar.share_metric),
-        ("subspace_metric", sidecar.subspace_metric),
         ("resolved_fit_mode", sidecar.resolved_fit_mode),
         ("topology_winner", sidecar.topology_winner),
         ("topology_candidates", sidecar.topology_candidates),

drowse/store/manifold_schema.py

@@ -619,8 +619,6 @@ class ManifoldSidecar:
     fit_policy_version: Optional[int] = None
     # ``mahalanobis`` on an ordinary fit; a monopolar fold labels its raw-δ̂
     # basis ``euclidean`` (a basis-selection label, not a metric fallback).
-    share_metric: Optional[str] = None
-    subspace_metric: Optional[str] = None
     # Curved-fit summaries: GCV smoothing ``{L: {lambda, edf, gcv}}`` and the
     # fuzzy-manifold tube ``{L: {sigma_mean, sigma_min, sigma_max, lambda}}``.
     rbf_smoothing_per_layer: dict[str, Any] = field(default_factory=dict)
@@ -839,11 +837,17 @@ def _non_negative_int(value: Any) -> bool:
 def validate_sidecar_payload(
     data: Any, *, location: str = "manifold sidecar",
 ) -> dict[str, Any]:
-    """Validate every field of the exact current fitted-sidecar contract."""
+    """Validate every field of the current fitted-sidecar contract.
+
+    Keys the contract no longer names (retired provenance) are dropped, so
+    removing a field never strands an existing fit; every current field must
+    be present.
+    """
     from drowse.geometry.domains import validate_domain_spec
 
-    if not isinstance(data, dict) or set(data) != SIDECAR_FIELDS:
-        raise ManifoldFormatError(f"{location} does not match the current exact schema")
+    if not isinstance(data, dict) or not SIDECAR_FIELDS <= set(data):
+        raise ManifoldFormatError(f"{location} does not match the current schema")
+    data = {key: data[key] for key in SIDECAR_FIELDS}
     validate_manifold_format_version(data["format_version"], location=location)
     try:
         validate_domain_spec(data["domain"])
@@ -881,7 +885,7 @@ def validate_sidecar_payload(
     nullable_strings = (
         "nodes_sha256", "sae_release", "sae_revision", "sae_fingerprint",
         "model_fingerprint", "model_source_fingerprint", "capture_sha256",
-        "share_metric", "subspace_metric", "resolved_fit_mode",
+        "resolved_fit_mode",
         "topology_winner", "bake_policy", "source_model_id",
         "source_model_fingerprint",
     )

pyproject.toml

@@ -89,7 +89,6 @@ addopts = ["--strict-markers"]
 xfail_strict = true
 markers = [
     "gpu: requires a CUDA or MPS backend (deselect with `-m \"not gpu\"`)",
-    "gpu_gemma: requires explicit gated Gemma model and artifact access",
     "perf: wall-clock throughput test; skipped under xdist, run with `pytest -m perf -p no:xdist`",
 ]
 filterwarnings = [