[a9lim]

Python goldens for drowse.ai's hook ABI v5 (units, token anatomy, the mid site and the S lens on WebGPU), plus the small contract and engine changes the browser parity plan needs.

Goldens

tests/support/browser_units.py owns the units-tiny-models-v1 fixture: gemma3_text and qwen3, 4 layers, d = 64, d_ff = 96, 4/2 heads × 16. Every weight comes from a portable 32-bit integer generator, so drowse.ai can rebuild the tensors bit for bit, and every linear and the embedding is q4f32_1-exact (power-of-two group scales with a ±7 in each group). The module has one build_<golden> per file and write_goldens, which enforces the validator and the size caps. Regenerate with PYTHONPATH=. python -m tests.support.browser_units --write.

New files under tests/fixtures/browser/. There is one model file, plus one file per arch for each of: block-sites, unit-stats, unit-ablation, unit-probes, unit-activity, lens-reads, slens, token-anatomy, attention-patterns, unit-reads, unit-artifacts and neuron-push. That is 25 files, 1.85 MB in total, the largest 191 KB. The caps are 256 KiB per file and 2 MiB overall. No existing golden changes.

unit-probes records one entry per forward (prefill and decode steps) from the unit tap's ring, with a gated case on a gate-only unit and the gate scalars the decode loop saw. unit-activity covers generated, sink, sink-population, authored (a user turn's prompt capture), two-spans, all-excluded, excluded-and-zero and segmented. lens-reads includes S direction reads (uncentred) and a session with no S lens. slens records the lowered 0.3 slens/<word> push, and token-anatomy carries each case's intermediates.

Every value comes from drowse's own machinery: tap_token, the real steering scope and decode loop, UnitHook/UnitOp, NodeReplayer (unit_activity, token_anatomy, _activity for the hand-built forwards), UnitReader, the lens instrument's reads and the S embedding statistics.

test_units_golden_matches_python rebuilds each golden from a fresh fixture and compares it within the file's own tolerance block. The model file must regenerate byte for byte, and its q4 round trip is asserted.

The validator refuses non-finite values, a sink row that moves under unit ablation, an ablation that moves no logit, a saturated attention row, and every top-k boundary without a 1e-3 margin. Each ranked list (lens reads, unit reads, unitTop hits, footprint top-k, the anatomy's unit lists and pattern heads) is read again one deeper: a changed returned set or a k-th entry that ties the best excluded refuses, and a read moves to another k (recorded in the file) until its boundaries clear. Probability lists rank by ln p, the logit gap their order rests on. unitFind records each greedy pick with its runner-up and refuses a pick without the margin. Every greedy decode pick must be decisive, generated nodes decode distinct tokens (greedy with a presence penalty), and unit activity refuses a listed or traced unit whose peak row has no margin (outside the declared excluded-and-zero case) and a traced neuron that keeps one sign.

Goldens hold across platforms: the python tolerance is the library's (1e-5 + 1e-4|x|). When writing, every margin also clears twice that tolerance, and the parity test rebuilds on the file's recorded choices (each read's k, a case's top, the anatomy's knobs, the gate thresholds; build_golden(..., replay=)), so a rebuild on another platform never re-decides near a threshold.

A near tie inside a list does not refuse: it is a run in declared_ties ({path, key, groups}). The comparison matches a run's entries by identity and compares every field (ids, strings, values), with order kept between runs. Unit-activity peak rows without a margin are skipped by index, and float16 populations allow one ulp.

Engine and contract

• contracts.ts: StructuredHookProfileId gains standard-v4 and standard-v4-residual. standard-v1–v3 stay until drowse.ai's libraries are recompiled. CatalogInstrumentPack.kind is now core | jlens | slens | units | sae.
• UnitReader.population_artifacts(layer) returns, per layer, what the browser ships offline: the population scalars (R rho/lse/lse_neg, S norm/footprint_std), the contrast statistics (S fires pooled), the heads' OV contrast, the head-norm mean derivatives and the whitened write norms. null_basis() and head_kept_ids() are now public, so drowse.ai's pack builder doesn't reach into privates.
• _norm_spec now judges each convention relative to its gain. When both fit (every gain large), the module itself, called with a zero weight through torch.func.functional_call, answers the offset; the closer median is only the fallback. Before this, a bf16 module with gains of several hundred (gemma-3-270m's MLP output norms, up to about 1,650) failed the absolute offset gap even though 1 + w was clearly the better fit. The regression tests (bf16 gemma3_text 1 + w and gemma4_text plain, large gains, and twelve all-large draws per arch) fail before this change.
• --dtype float32|bfloat16|float16 on every model-loading verb. It reaches from_pretrained(dtype=) and the model_source_fingerprint that the lens-fit and manifold no-op proofs key on, so a bf16 fit is never proven as an fp32 one. The loaded model's stamp already carried the parameter dtype.

Checks

• pytest -q -m "not gpu and not perf" -n 8: 6022 passed, 7 skipped
• ruff check .: clean
• pyright: 0 errors (the two remaining warnings are pre-existing)
• scripts/generate_webui_types.py --check: current
• cd webui && npm ci && npm run check && npm run build: Node 24.22.1, passes, no dist drift
• GitHub CI (Linux, Python 3.11 to 3.13): green