From 1692b16f6e8d96fc3414492a6e558fcf84935039 Mon Sep 17 00:00:00 2001 From: rjperry36 Date: Fri, 17 Jul 2026 11:59:27 +0100 Subject: [PATCH] fix(sensing-server): canonicalize calibration frames to 56-tone grid (real HT40 fix) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Follow-up to the calibration deadlock fix. With the status gate unstuck, maybe_feed_calibration reached feed_calibration, but a real ESP32 HT40 node streams 128-wide amplitude frames while the single-link FieldModel is the canonical 56-tone grid. LinkStats::update returned DimensionMismatch, feed_calibration bubbled it, and maybe_feed_calibration swallowed it at debug level — so frame_count stayed pinned at 0 on live hardware (presence/vitals, which read the global history, were unaffected). Resample each frame onto the model's canonical 56-tone grid via HardwareNormalizer::resample_to_canonical before feeding — the same length-only canonicalization the multistatic fusion path uses (#1170). Pinned by maybe_feed_calibration_resamples_wide_frames_and_accumulates (128-wide -> Collecting + count 1). sensing-server bin: 173 passed, 0 failed. Co-Authored-By: claude-flow --- CHANGELOG.md | 1 + .../src/field_bridge.rs | 47 ++++++++++++++++++- 2 files changed, 46 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 6cabed94..c44ddb5b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,6 +16,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 - **`/api/v1/stream/pose` WebSocket reachable with `RUVIEW_API_TOKEN` set + dashboard bearer-token field** (#1310, PR #1313, @erichkusuki). Browsers cannot attach an `Authorization` header to a WS upgrade, so the Live Demo pose stream always failed when auth was on; the path is now on a narrow exact-match exemption list (mirrors `/ws/sensing`), with a regression test pinning that the exemption doesn't leak to other `/api/v1/*` paths. The QuickSettings panel gains an "API Access" field storing the bearer token in `localStorage`; the token is applied at `api.service.js` module load so the very first request carries it. - **Display-less DevKitC-1 boards: `sdkconfig.defaults.devkitc` build overlay** (#1308, PR #1311, @erichkusuki). The ADR-045 runtime display probe false-positives on stock ESP32-S3-DevKitC-1 (floating QSPI pins), which silently skipped the RuView#893 MGMT+DATA CSI upgrade and collapsed CSI yield to 0 pps. The overlay compiles display support out (`has_display` constant-false). Also fixes stale `espressif/idf:v5.2` README references to v5.4 (source requires `esp_driver_uart`, IDF ≥5.3). Hardware-verified on 2× DevKitC-1-N16R8 (0 → 40–45 pps). - **ADR-263/264/265 implemented — the RuView npm surface fixed end-to-end (`@ruvnet/ruview@0.2.0`, `@ruvnet/rvagent@0.2.0`, `@ruv/ruview-cli`).** Harness (ADR-263 O1–O9): `claim-check` now **fails closed** on empty input (CLI exit 2 + `empty_text` tool error); the MCP stdio server dispatches `tools/call` asynchronously over promise-based `spawn` — `ping` answers while a long `verify`/`calibrate` runs (pinned by a new e2e test that runs a 3 s fake proof and asserts sub-second ping); the two `optionalDependencies` are gone so a cold `npx` installs exactly 1 package (MEASURED: was 4 packages / 620 kB / 71 files, `npm i` in a clean prefix); child output is captured as bounded rolling tails (no more 1 MiB `maxBuffer` kills); `node_monitor` passes the port via `sys.argv` instead of splicing it into `python -c` source; the MCP `serverInfo.version` reads package.json; `.claude/skills/*/SKILL.md` are generated from `skills/*.md` by a `prepack` sync script (byte-equality pinned by test); `which()` is a memoized dep-free PATH scan; tools are underscore-canonical (`ruview_claim_check`, …) with the dotted names accepted as call-time aliases, plus `resources/list`/`prompts/list` stubs; the guardrail's `METRIC_TERMS` matching is precision-fixed (word-boundary `map`/`f1`/`auc`/`iou`, code-span + label scrubbing, quantitative-claims-only) — ADR-263/264/265 and both package READMEs now PASS `claim-check` while real untagged claims still flag. 30/30 tests (MEASURED, `node --test`). rvagent (ADR-264 O1–O9): `exports` fixed (types-first, the never-built `dist/index.cjs` `require` target removed — verified broken in the published 0.1.0 tarball); tarball is map-free (127,704 B unpacked / 46 files / 0 maps — MEASURED, `npm pack --dry-run`, down from 188 kB with 44 maps); the Streamable HTTP transport is **actually wired** behind `RVAGENT_HTTP_PORT` with one transport + one MCP server per session (`mcp-session-id` routing), a 1 MiB body cap (413), and a port-aware localhost origin gate — the "dual-transport" description is now true; tools renamed to underscore-canonical with dotted router aliases; ONE Zod validation gate per call with the advertised JSON Schema generated from the same Zod source (`zod-to-json-schema`); `train_count` closes its log fds (was leaking 2/job) and persists job records to `/.json` so `job_status` survives restarts, with bounded log-tail reads; `detectCogBinary` actually probes its candidate paths; version reads package.json; `@types/express` dropped, `@types/jest` aligned to jest 29; README rewritten to match reality (no phantom `stdio`/`http`/`policy grant` subcommands; unimplemented ADR-124 catalog tools labeled roadmap). 99/99 jest tests (MEASURED); stdio handshake + HTTP session flow + 403/400/404/413 gates smoke-tested live. CLI: bin renamed `ruview-cli` (the `ruview` bin belongs to `@ruvnet/ruview`, ADR-265 D4), version single-sourced. Distribution (ADR-265 D1–D4): new `npm-packages.yml` (3-package × Node 20/22 matrix: tests, version-literal grep gate, pack-content/size gate, tarball-install smoke test incl. the fail-closed claim-check and an ESM-import probe that would have caught the broken `require` export, README claim-check) and `ruview-npm-release.yml` (publish from CI only, `npm publish --provenance`); `ci.yml` NODE_VERSION 18→20. +- **Empty-room field-model calibration collected nothing on real HT40 nodes — raw 128-wide frames rejected by the 56-tone model (follow-up to the deadlock fix below).** Once the status-gate deadlock was fixed, `maybe_feed_calibration` reached `feed_calibration`, but a real ESP32 HT40 node streams 128-wide amplitude vectors while the single-link `FieldModel` is the canonical 56-tone grid — so `LinkStats::update` returned `DimensionMismatch`, `feed_calibration` bubbled the error, and `maybe_feed_calibration` swallowed it at `debug` level. Net effect: `frame_count` stayed pinned at 0 on live hardware even though presence/motion/vitals (which read the global history) worked fine. Fixed by resampling each frame onto the model's canonical 56-tone grid via `HardwareNormalizer::resample_to_canonical` before feeding — the same length-only canonicalization the multistatic fusion path uses (#1170). Pinned by `field_bridge::maybe_feed_calibration_resamples_wide_frames_and_accumulates` (128-wide frame → Collecting + count 1; fails on old code). Verified live on the ESP32-S3 deployment. - **Empty-room field-model calibration could never start — `/api/v1/calibration/*` was a dead endpoint (frame count pinned at 0).** `POST /calibration/start` creates the `FieldModel` in `Uncalibrated`, but the per-frame server feed `field_bridge::maybe_feed_calibration` only fed observations while the model was **already** `Collecting` — and the *only* thing that sets `Collecting` is `feed_calibration` itself (on its first fed frame). The two gates deadlocked: nothing ever fed the first frame, so `calibration_frame_count` stayed 0, `status` never left `Uncalibrated`, and the SVD room eigenstructure (eigenvalue-based person counting / localization) could never calibrate — observed live as `{"status":"Uncalibrated","frame_count":0}` never advancing on a streaming ESP32 node. Fixed the guard to feed while `Uncalibrated | Collecting` so the first frame flips the model to `Collecting` and the count advances. Also made `calibration_stop` return a structured `{success:false, frame_count, frames_needed}` (with a new `FieldModel::min_calibration_frames()` accessor) instead of an opaque 500 when finalized before enough empty-room frames accumulate. Pinned by `field_bridge::maybe_feed_calibration_advances_uncalibrated_to_collecting` (asserts `Uncalibrated → Collecting` + count 0 → 1 → 2; fails on old code). Presence/motion/vitals were unaffected — they use the separate auto rolling baseline, not the field model. - **Multistatic fusion never ran on a mixed-mode ESP32 mesh — live bridge fed raw, un-canonicalized per-node CSI to the fuser (#1170).** `node_frame_from_state` (`multistatic_bridge.rs`) wrapped each node's **raw** amplitude vector (HT20 ≈ 64 bins, HT40 ≈ 128/192) into a struct *named* `CanonicalCsiFrame` without ever resampling, so `MultistaticFuser::fuse` tripped `DimensionMismatch` on every cycle, silently fell back to per-node sum/dedup, and spun `total_engine_errors` unbounded. Added `HardwareNormalizer::resample_to_canonical` (resample-only, **no z-score** — preserves the amplitude scale the person-score's `variance/mean²` relies on) and run every node frame through it onto the canonical 56-tone grid before fusion. Heterogeneous meshes now fuse instead of erroring. Pinned by `heterogeneous_node_counts_canonicalize_and_fuse` (mixed 64/192 → fuses), `resample_to_canonical_is_length_only_no_zscore`, and an updated `test_node_frame_conversion`; the pre-existing `engine_bridge::observe_cycle_counts_engine_errors` was retargeted to force a `TimestampMismatch` (its old 56-vs-30 setup now canonicalizes cleanly). `wifi-densepose-signal` 501 / `wifi-densepose-sensing-server` 677 tests, 0 failed. - **`csi_fps_ema` reported the CSI frame rate 40–840× too high under bursty UDP delivery (#1180).** `update_csi_fps_ema` only rejected deltas `≤ 0` or `≥ 1 s`, so a 36 µs intra-burst arrival delta yielded `1/dt ≈ 27 kHz` straight into the EMA — the metric measured server arrival jitter, not the node's ~40 fps production rate. Added a `MIN_PLAUSIBLE_CSI_DT_SEC = 0.005` floor (derived from the firmware's 50 fps `CSI_MIN_SEND_INTERVAL_US` ceiling, ×4 slack) and made `observe_csi_frame_arrival` keep its anchor across sub-floor bursts so the next genuine inter-frame gap measures true cadence. Pinned by `subms_burst_delta_rejected`, `burst_interleaved_with_nominal_stays_in_band`, and `observe_csi_frame_arrival_ignores_subms_bursts`. diff --git a/v2/crates/wifi-densepose-sensing-server/src/field_bridge.rs b/v2/crates/wifi-densepose-sensing-server/src/field_bridge.rs index f77ed93c..5b7f588f 100644 --- a/v2/crates/wifi-densepose-sensing-server/src/field_bridge.rs +++ b/v2/crates/wifi-densepose-sensing-server/src/field_bridge.rs @@ -7,12 +7,22 @@ //! score-based heuristic in `score_to_person_count`. use std::collections::VecDeque; +use std::sync::LazyLock; +use wifi_densepose_signal::hardware_norm::HardwareNormalizer; use wifi_densepose_signal::ruvsense::field_model::{ CalibrationStatus, FieldModel, FieldModelConfig, }; use super::score_to_person_count; +/// Length-only canonicalizer for calibration frames (issue #1170 pattern, +/// shared with `multistatic_bridge`). Raw ESP32 amplitudes arrive at the +/// hardware's native width (HT20 ≈ 64, HT40 ≈ 128/192); the FieldModel is +/// configured for the canonical 56-tone grid, and `feed_calibration` rejects +/// any other width with `DimensionMismatch`. Resampling here (default 56) +/// lets real HT40 nodes actually calibrate instead of silently feeding nothing. +static CALIB_NORMALIZER: LazyLock = LazyLock::new(HardwareNormalizer::new); + /// Number of recent frames to feed into perturbation extraction. const OCCUPANCY_WINDOW: usize = 50; @@ -113,8 +123,13 @@ pub fn maybe_feed_calibration(field: &mut FieldModel, frame_history: &VecDeque> = VecDeque::new(); + history.push_back(wide); + + maybe_feed_calibration(&mut field, &history); + + assert_eq!( + field.status(), + CalibrationStatus::Collecting, + "128-wide frame must resample to 56 and be accepted" + ); + assert_eq!( + field.calibration_frame_count(), + 1, + "wide frame must accumulate, not be silently dropped" + ); + } }