a9lim · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -28,75 +28,26 @@ 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]"
       - run: pyright
-      # The dashboard's REST types are generated from the FastAPI app's
-      # OpenAPI schema, so an unmirrored response-shape change has to fail
-      # the build rather than a panel.  It lives in this job rather than in
-      # ``webui`` because the generator imports the package: this is the job
-      # that already has it installed, and duplicating the install would
-      # double CI's heaviest step for one diff.  ``webui`` keeps the
-      # sibling parity check for the committed bundle.
-      - name: Verify the generated dashboard types are current
-        run: python scripts/generate_webui_types.py --check
 
-  test-matrix:
+  test:
     runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.11", "3.12", "3.13"]
     steps:
       - uses: actions/checkout@v7
       - uses: actions/setup-python@v7
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.13"
+          cache: pip
+          cache-dependency-path: pyproject.toml
       - run: python -m pip install --upgrade pip
       - run: pip install -e ".[dev]"
-      # Coverage is informational only — no --cov-fail-under, and --cov stays out
-      # of pyproject addopts so local runs without pytest-cov keep working.
-      - run: pytest -q -m "not gpu" --cov=drowse --cov-report=term-missing
-
-  # Stable aggregator so branch protection can require a single ``test`` context
-  # instead of every ``test-matrix (x.y)`` leg by name. ``if: always()`` makes the
-  # gate run even when a leg fails, and the explicit result check turns a failed or
-  # skipped matrix into a failed gate (a skipped required check would otherwise let
-  # a PR merge).
-  test:
-    runs-on: ubuntu-latest
-    needs: [test-matrix]
-    if: always()
-    steps:
-      - name: Verify the test matrix succeeded
-        run: |
-          if [ "${{ needs.test-matrix.result }}" != "success" ]; then
-            echo "test-matrix did not succeed: ${{ needs.test-matrix.result }}"
-            exit 1
-          fi
-
-  build:
-    # Validates the PEP 621 dynamic-version setup in pyproject.toml: confirms
-    # ``python -m build`` resolves ``drowse.__version__`` via AST parse without
-    # executing the module (which would import torch and break the build).
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v7
-      - uses: actions/setup-python@v7
-        with:
-          python-version: "3.12"
-      - run: python -m pip install --upgrade pip build
-      - name: Seed a stale dashboard build cache
-        run: python scripts/check_package_isolation.py seed
-      - run: python -m build --sdist --wheel
-      - name: Verify wheel dashboard files match the source distribution
-        run: python scripts/check_package_isolation.py check
-      - name: Verify wheel imports and version round-trips
-        run: |
-          pip install dist/drowse_ai-*.whl
-          python -c "import drowse; from importlib.metadata import version; assert version('drowse.ai') == drowse.__version__; print('version:', drowse.__version__)"
+      - run: pytest -q -m "not gpu" -n auto
 
   webui:
     runs-on: ubuntu-latest

diff --git a/AGENTS.md b/AGENTS.md
@@ -117,7 +117,7 @@ deprecation shims or second spellings.
 
 ## Package map and layer order
 
-`tests/invariants/test_layering.py` enforces this order, lowest first. A
+The order runs lowest first. A
 module may import its own layer or any lower one. Every import counts:
 module-level, function-level and `if TYPE_CHECKING:`. A lower layer that needs
 a higher object types it against a narrow `Protocol` of its own
@@ -306,9 +306,8 @@ 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 types `--check`, the CPU
-suite on Python 3.11–3.13, the package-isolation build and the webui build
-with a dist-drift check. Details: `tests/AGENTS.md`.
+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`.
 
 ## State and environment
 

diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
@@ -72,7 +72,7 @@ next steered forward.
 | Conversation tree, recipes, replay, transcripts | `loom/` |
 | Protocols, dashboard mount, CLI | `server/`, `web/`, `cli/` |
 
-Dependencies point downward only (`tests/invariants/test_layering.py`). A
+Dependencies point downward only. A
 collaborator gets its dependencies through its constructor; nothing outside
 `drowse/session/` reads a session private. That rule is what lets each
 collaborator be tested without a session and lets the browser runtime replace

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -28,7 +28,7 @@ pytest -q tests/gpu/test_smoke.py   # real-model smoke tests
 ```
 
 Tests marked `gpu` need CUDA or Apple Silicon MPS and may download model weights.
-CI runs the non-GPU suite on Python 3.11, 3.12, and 3.13.
+CI runs the non-GPU suite on Python 3.13.
 
 ## Static checks
 

diff --git a/drowse/cli/AGENTS.md b/drowse/cli/AGENTS.md
@@ -4,7 +4,7 @@ The `drowse` console script (`drowse.cli.main:main`; `python -m drowse` is the
 same entry point): nine verb groups — `serve`, `manifold`, `pack`, `config`,
 `experiment`, `template`, `lens`, `sae`, `unit` — as thin frontends over
 `DrowseSession` and the `store`/`fit`/`lens`/`sae`/`units` packages. The CLI is in the top layer beside
-`server`, `web` and `notebook` (`tests/invariants/test_layering.py`): it may
+`server`, `web` and `notebook`: it may
 import any lower layer and never reaches a session private (`session._x`).
 
 ## Module map

diff --git a/drowse/fit/AGENTS.md b/drowse/fit/AGENTS.md
@@ -4,7 +4,7 @@ Corpora in, fitted manifold tensors out: the one extraction pipeline, corpus
 generation, the neutral-activation capture behind the whitener, the offline
 bake, cross-model alignment orchestration, and the naturalness eval. `fit`
 sits above `generation` and below `loom`/`session`
-(`tests/invariants/test_layering.py`), so it never imports the session; hosts
+so it never imports the session; hosts
 are Protocols (`ModelHandle`, `NaturalnessHost`) that `DrowseSession`
 satisfies. The math it calls is in `geometry/AGENTS.md` (Fisher basis, frame,
 share, discover coords, topology, RBF, σ-field, fold). The folder format,

diff --git a/drowse/generation/AGENTS.md b/drowse/generation/AGENTS.md
@@ -8,8 +8,8 @@ routing, the transaction) is `session/engine.py`; the multi-row runner is
 `session/batch.py` (see `drowse/session/AGENTS.md`).
 
 Layer: above `readout`, `lens`, `sae`, `steering.composer`/`hooks`; below `fit`,
-`loom`, `session`. `tests/invariants/test_layering.py` counts `TYPE_CHECKING`
-imports too, so collaborators here type against a narrow `Protocol`
+`loom`, `session`. `TYPE_CHECKING`
+imports count too, so collaborators here type against a narrow `Protocol`
 (`render.ConversationHistory`) or a lower-layer class.
 
 | Module | Owns |

diff --git a/drowse/geometry/AGENTS.md b/drowse/geometry/AGENTS.md
@@ -7,7 +7,7 @@ whitener, the direction view (fold, `Profile`, GGUF) and cross-model transfer.
 
 Pure tensor math. Geometry sits directly above `errors`/`naming`/`events` and
 `base` in the layer order and imports nothing else from drowse
-(`tests/invariants/test_layering.py`; `drowse.__version__` is exempt). The
+(`drowse.__version__` is exempt). The
 manifold tensor codec is `store/manifold_tensors.py`; capture and fit
 orchestration are `fit/` (`pipeline.py`, `node_stats.py`,
 `capture.py::compute_dls_axes`); the gains that scale the kernel are in

diff --git a/drowse/lens/AGENTS.md b/drowse/lens/AGENTS.md
@@ -13,7 +13,7 @@ atom `K_l ẽ_v` and the `tokens-in` decomposition. `read.py` reads any vector
 at a place through both, by the four-lens rule. Inside a session the face is
 `session.lens` (`LensInstrument`). The package sits beside `sae`, above `readout` and below
 `steering.composer`, and imports only `errors`, `base`, `model` and `readout`
-(`tests/invariants/test_layering.py`, which counts `TYPE_CHECKING` imports).
+(`TYPE_CHECKING` imports included).
 
 | Module | Owns |
 |---|---|

diff --git a/drowse/loom/AGENTS.md b/drowse/loom/AGENTS.md
@@ -28,8 +28,8 @@ recipes; the verbs, `restore_tree`, the joint-logprob cache),
 
 ## Layering
 
-`loom` sits above `fit` and below `session`. `tests/invariants/test_layering.py`
-enforces this, `TYPE_CHECKING` imports included. Lower layers never import
+`loom` sits above `fit` and below `session`. `TYPE_CHECKING`
+imports count too. Lower layers never import
 loom: `generation/render.py` types the tree as `ConversationHistory`, and
 `generation/finalize.py` takes `tree: Any`.
 

diff --git a/drowse/model/AGENTS.md b/drowse/model/AGENTS.md
@@ -5,7 +5,7 @@ loading and identity, where the residual blocks are and what they compute
 (`BlockSpec`), unit directions, the logits path, layer selectors, token spans, chat-template facts, role substitution, the scene
 stitcher, and compile/StaticCache plumbing. The package imports only
 `errors`/`naming`/`events`, `base` and `geometry`
-(`tests/invariants/test_layering.py`) and holds no session state: every
+and holds no session state: every
 function takes a model, tokenizer or config.
 Render *policy* (which path a conversation takes) lives in
 `generation/decode.py::build_chat_input` and `generation/render.py`; this

diff --git a/drowse/readout/AGENTS.md b/drowse/readout/AGENTS.md
@@ -35,7 +35,7 @@ other three families (`drowse/units/AGENTS.md`);
 ## Layering
 
 `readout` sits above `steering.{grammar,terms,triggers}` and below
-`lens`/`sae`/`units` (`tests/invariants/test_layering.py`; `TYPE_CHECKING`
+`lens`/`sae`/`units` (`TYPE_CHECKING`
 imports count). It must not import `lens`, `sae`, `units`,
 `steering.composer`/`hooks`, `generation` or
 anything above; collaborators from those layers are typed against the

diff --git a/drowse/sae/AGENTS.md b/drowse/sae/AGENTS.md
@@ -9,7 +9,7 @@ local training, and the `sae` read instrument (`session.sae`,
 
 `sae` shares a layer with `lens` (above `readout`, below `steering.composer`):
 it never imports `generation`, `fit`, `loom` or `session`, type-only imports
-included (`tests/invariants/test_layering.py`). `DrowseSession.__init__` wires
+included. `DrowseSession.__init__` wires
 every dependency explicitly.
 
 | Module | Owns |

diff --git a/drowse/server/AGENTS.md b/drowse/server/AGENTS.md
@@ -30,9 +30,8 @@ an API path.
 ## Rules for editing
 
 - `server/` is in the top layer with `cli`, `web`, `notebook`. It never reads
-  a session-private attribute (`tests/invariants/test_layering.py`), and no
-  server module imports a sibling's `_private` helper
-  (`tests/invariants/test_boundary_guard.py`); shared helpers go in
+  a session-private attribute, and no server module imports a sibling's
+  `_private` helper; shared helpers go in
   `runtime.py`, `errors.py`, `operations.py`, or `request_helpers.py`.
 - Native request bodies subclass `NativeRequest` (`extra="forbid"`), nested
   objects included. Ollama bodies are `extra="allow"` (clients send fields

diff --git a/drowse/session/AGENTS.md b/drowse/session/AGENTS.md
@@ -9,7 +9,7 @@ collaborator with explicit constructor dependencies, not into session state.
 Outside this package nothing reads a session private (`session._x`, `sess._x`,
 `<x>.session._x`, `self._session._x`); modules inside it may (the engine calls
 `_start_loom_assistant`, `_plan_authored_prompt_capture`, `_steering_composer`).
-`tests/invariants/test_layering.py` enforces this and pins `session` directly
+`session` sits directly
 below `cli`/`server`/`web`/`notebook`. Two collaborators take the session as a
 Protocol-typed handle, so keep those properties stable: `Extraction`
 (`fit.pipeline.ModelHandle`: `model`, `tokenizer`, `model_id`, `device`,

diff --git a/drowse/steering/AGENTS.md b/drowse/steering/AGENTS.md
@@ -15,7 +15,7 @@ injected, where, when, and how strongly.
 | `composer.py` | `SteeringComposer` (stack, scopes, resolution, composition, plan cache, stack queries, `term_layers`), `SteeringScope`, `TermLayers`, `resolve_scope_role`, `select_term_layers`, `affine_manifold_push`, the `ArtifactSource` / `NeutralSource` protocols |
 | `hooks.py` | `SteeringPlan`, `_lower_affine_subspaces`, `AffineOp`, `SteeringHook`, `UnitTerm` / `UnitOp` / `UnitHook` (unit ablation), `SteeringManager`, `install_persistent_offset_hooks`, the gain constants |
 
-Layering (`tests/invariants/test_layering.py`): `grammar`/`terms`/`triggers`
+Layering: `grammar`/`terms`/`triggers`
 sit above `store` (parsing resolves bare atoms against installed manifolds via
 `store.selectors`) and below `readout`; `composer`/`hooks` sit above
 `lens`/`sae` and below `generation`. Lower layers that must refuse work under

diff --git a/drowse/store/AGENTS.md b/drowse/store/AGENTS.md
@@ -7,8 +7,8 @@ alignment caches.
 
 Layer: above `model`, below `steering.grammar` (`parse_expr` resolves bare atoms
 through `store.selectors`). Store imports only `errors`/`naming`/`events`, `base`,
-`geometry`, `model` and itself; `tests/invariants/test_layering.py` counts
-function-level and `TYPE_CHECKING` imports too. No barrel module: import from the
+`geometry`, `model` and itself; function-level
+and `TYPE_CHECKING` imports count too. No barrel module: import from the
 submodule. This package owns formats and folder primitives; the fit's folder
 transactions (authoring snapshot, compare-and-swap publish) are `fit/publish.py`,
 the weight-free fit no-op proof `fit/preflight.py`, the capture cache

diff --git a/drowse/units/AGENTS.md b/drowse/units/AGENTS.md
@@ -17,7 +17,7 @@ with the other steering hooks in `steering/hooks.py`.
 | `reads.py` | `UnitReader` (static reads, `top`, `find`, `stats`), the results (`NeuronRead`/`NeuronPoleRead`, `HeadRead`, `PlaceRead`/`UnitHalf`, `Firing`, `TopResult`/`TopHit`, `FindResult`), `ContrastCache`/`ContrastStats`, `own_token_z`, `copy_score`, `logit_moments`, `token_logits` |
 | `catalog.py` | the neuron-lens catalog: `CATALOG_REPO`/`CATALOG_REVISION`/`CATALOG_MODELS`, `fetch_catalog`, `catalog_status`, `list_catalogs`, `remove_catalog`, `UnitCatalog` (rows, labels, `url`), `CatalogBinding` (the session's view), `unit_url`, `LABEL_CREDIT`, the errors |
 
-Layer 7, beside `lens` and `sae` (`tests/invariants/test_layering.py`): it
+Layer 7, beside `lens` and `sae`: it
 imports `model`, `store`, `readout`, `lens` and below, never
 `steering.composer`, `generation` or the session. The composer (layer 8)
 reaches it through its `UnitSource` protocol; the session wires everything

diff --git a/drowse/web/AGENTS.md b/drowse/web/AGENTS.md
@@ -77,29 +77,9 @@ npm run build        # vite build into ../drowse/web/dist + check-build-isolatio
 - The SAE description indexes in `lib/data/sae-descriptions/` ship on purpose (as
   `assets/gemma-3-{270m,1b,4b}-it.json`): they label local SAELens gemma-scope-2 releases
   offline. Regenerate them with `node scripts/build-sae-descriptions.mjs`.
-- CI's `build` job seeds a stale `build/lib` cache, then runs
-  `scripts/check_package_isolation.py check`. The wheel's dashboard files must equal the
-  sdist's and the source tree's.
 - `npm run check` runs `svelte-check`, `check-theme.mjs` (every referenced CSS custom
   property must be declared, here or in `@stdlogits/ledger`, so an old token name fails
-  here), `check-ledger.mjs` (the Ledger ratchet, below), `check-runtime-boundary.mjs`, the
-  `test:*` node suites and `check-test-coverage.mjs`. The last requires every
-  `scripts/*.test.mjs` to be reachable from `check`, so a new test file needs a line in
-  `package.json`.
-- `check-ledger.mjs` counts, per file, the departures from Ledger: borders; outlines outside
-  a `:focus`/`:focus-visible` rule; shadows other than a marked token's 2px same-colour spread
-  (the tip's `--float` is Ledger's own, so any use here counts); monospace; capitals;
-  letter-spaced text; easing off the one curve; animations that loop; colours outside the
-  tokens; colour mixes outside oklch (the accent's) and oklab (a grey's: Chrome renders a
-  near-grey mixed in oklch with a red cast); and text coloured by the depth accent (`--tint`,
-  `--deep`, `--late`, or a component's own accent property). It reads `style:prop={…}`
-  directives as `prop: …` declarations. It fails if any count differs from
-  `webui/ledger-ratchet.json`, which holds every count at zero: the bridge that once mapped
-  the old token names onto Ledger is gone, and with it the ratchet's bridge category. A count
-  that rises fails; one that falls asks for `node scripts/check-ledger.mjs --write`, which
-  lowers the ratchet and never raises it. A category the ratchet does not record yet fails
-  the check until `--write` takes today's count as its first ceiling. `--cut` (the 6.0 cut)
-  requires every count at zero.
+  here), `check-runtime-boundary.mjs` and the `test:*` node suites.
 
 ## Generated REST types (`lib/types.gen.ts`)
 
@@ -674,7 +654,7 @@ token).
 the theme preflight), the 404 for a missing asset or any file the bundle lacks, the SPA
 fallback, traversal, a public shell under an
 API key, `web=False`, and an empty dist. `tests/invariants/test_webui_types_parity.py`
-gates the generated types, and `test_layering` places `web` in the top layer.
+gates the generated types.
 
 ## Deliberately absent