PR: ruvnet/RuView#1387 Announcement page: https://claude.ai/code/artifact/0d3a1b19-b3b0-4a69-be73-183c9e42ce59?org=edd34d62-ff8b-4226-aceb-42b4cd07b607 Interactive capabilities showcase: https://claude.ai/code/artifact/7dfda175-a64f-4a2a-b01f-f3469d1a85d5?org=edd34d62-ff8b-4226-aceb-42b4cd07b607
A compiled, PyO3-bound wifi-densepose pip package for passive WiFi-CSI sensing now ships three opt-in SOTA extras, each proven bit-identical to the native Rust core via a SHA-256 parity harness.
- AETHER — 128-dim contrastive CSI embeddings for room fingerprinting / re-identification, now with an optional trained-weight loader.
pip install wifi-densepose[aether] - MERIDIAN — hardware-invariant CSI normalization (ESP32/Intel/Atheros), AP-geometry encoding, few-shot LoRA room calibration.
pip install wifi-densepose[meridian] - MAT — CSI-based disaster-survivor detection with START-protocol triage.
pip install wifi-densepose[mat] - A real production bug fixed along the way: same-second training runs were silently overwriting each other's model files (
.rvffilename collision) — found via CI, root-caused, fixed, and pinned with a regression test. - The
pip-releaseCI pipeline's root failure (a stale, 403-ing PyPI token) diagnosed and fixed. A premature migration to PyPI OIDC Trusted Publishing was caught by a pre-existing regression guard (it would have left publishing broken pending a manual registration step) and correctly reverted. - The dashboard's "Start Training" button — previously a silent no-op — now drives a real trainer with live
/ws/train/progressstreaming. archive/v1(the old pure-Python implementation) formally deprecated with honest model-weight labeling.python/is now gated in a real, blocking CI workflow (python-ci.yml) instead of only being checked at release time.
| Default wheel | 279 KB |
[sota] combined wheel |
2.2 MB (budget: ≤5 MB) |
| Tests passing | 218 |
| Parity fixtures | 6/6 (binding == native Rust, SHA-256 gated) |
| Python support | 3.10–3.13, one abi3 wheel per platform |
- Live PyPI publish is blocked on registering a PyPI Trusted Publisher — a one-time, account-holder-only step on pypi.org. The wheel builds and passes every test; it just isn't live on PyPI yet.
- SOTA accuracy validation: the parity harness proves the Python bindings are bit-identical to the native Rust code — a real, strong guarantee — but it does not prove the native code hits its published accuracy bars (PCK@20, re-ID mAP, cross-domain MPJPE) against real labeled CSI. That needs a labeled dataset and eval harness that don't exist yet in the repo.
See PR #1387 for the full ADR series (ADR-184 through ADR-187) with commit-level detail.