Live report: skeleton straight but hands off the wrists. The app runs
Vision with a hardcoded .right orientation (ARBodySession.swift:261),
so hand coords live in an uprighted frame while ARKit skeleton2D lives
in the raw sensor frame. Hands get their own IPHONE_HANDS_ROTATE knob
(default none); VIDEO_ROTATE now only rotates the skeleton2D points.
SC handlers (edge detector, tick, OSCdefs, pose/hands calls) and web
UI (Poings section, sensibilite poing slider, fistprogress indicators)
removed. Python FIST_ENABLE=0 kill-switch was already in effect.
Old .matrix files with fistThresh key load without error (key silently
ignored). Test harness Q section removed from test_global_actions.scd.
Live report: video straight but skeleton rotated wrong. Flip and
90-degree rotation do not commute — the video is flipped THEN rotated
while points are rotated THEN display-flipped. Points now use the
mirror-conjugated mode (cw/ccw swapped under mirror), proven by a
composition test.
Portrait mode validated live: VIDEO_ROTATE now also rotates the
iPhone 2D landmarks (skeleton2D + Vision hands) at the decode layer so
overlays/panels/gestures stay aligned with the rotated video. ARKit
per-finger joints are masked from the wireframe (unreliable 2D, seen
as false hands); the display shows the filtered Vision hands instead.
Apply hand_plausible + HandPersistenceGate in _step_stab before
route_hands so gestures and /pose/hands features only see quality-
gated hands. Cache per-tick validated set (_tick_hands/_tick_chir/
_tick_mirror) so all three emitters share one acquisition. Gate
stepped at 30 Hz publisher cadence (audit R6). Non-protocol hands
(raw numpy arrays) pass through via try/except for backward compat.
The body-mesh triangles (solid cyan blob reported live) were the one
surface no filter covered: under iphone-usb persons_body is the ARKit
conversion with c=1.0, so garbage/stale joints painted full triangles.
Skip BODY_TRIANGLES whenever the ARKit topology is present (webcam
MediaPipe path unchanged); the body stays the freshness-gated ARKit
wireframe.
config.py: PINCH_EXT_MIN docs said "min frames for debounce"
— it is min extended fingers required to fire; fix wording.
HAND_PERSIST_FRAMES said "persist after disappearance" — it
is frames required before a hand is drawn; fix wording.
CLAUDE.md: MEDIAPIPE_DELEGATE default was "cpu" — real
default is "gpu" (config.py); fix the RC0.1+ table row.
Add missing PINCH_ENABLE / VIZ_HUD / CONCERT_MIRROR rows
to the main env-vars table.
iphone_usb_source.py: after stream ends, clear
persons_hands_iphone, persons_hands_chirality, and
persons_hands (when write_hands=True) under state.lock()
to prevent renderer display freeze.
test_iphone_usb_reconnect.py: add
test_hands_cleared_on_stream_drop to verify lists are
empty after simulated USB disconnect.
renderer.py: read ARKIT_FULL_SKELETON via VizConfig at
initWithState_ instead of module-level os.environ; store as
self._arkit_full, use getattr fallback for __new__ test bypass.
main.py: pass cfg.iphone_osc_port to IphoneOSCListener
instead of hardcoded 57128; log actual port.
hand_slots.py: import _finite/_coord/_clamp from hand_features
at module level (acyclic: hand_features imports hand_slots only
lazily inside step()); delete the duplicate copies.
User request live 2026-07-02: disable the fist gesture completely
(false held-fist actions during play). fist is forced to 0.0 at the
extractor unless FIST_ENABLE=1; continuous features unchanged. SC
held-fist actions (reload/next) go silent without touching SC.
iphone-usb path: camera decode + ARKit conversion is cheap (no
MediaPipe inference on Mac). Frame-based gesture calibrations
(debounce/hold/grace/persist) assume 30 Hz. Keep 18 Hz for the
webcam MediaPipe path (M5 CPU protection).
Tests: period formula + MultiWorker.__init__ signature check.
Read state.mirror_2d in _pick_hands and thread it through
_step_stab (route_hands mirror=) and _emit_hands
(HandFeatureExtractor.step mirror=) so the cx fallback slot
assignment matches the renderer panel layout on chirality dropout.
Add mirror kwarg to HandFeatureExtractor.step plumbed to route_hands.
Tests: extractor-level mirror flip + publisher _emit_hands mirror_2d.
Detection v3 review follow-ups: pin hand_persist_grace/near_off/
hold_frames defaults; use the module-level route_hands (restores the
monkeypatch seam, drops the dead local alias); document the grace
count-inheritance window in the gate docstring.
hand_slots module docstring still claimed inverted chirality (default
1) after the live A/B validated chirality correct (default 0); align
it. Pin the MESH_RIG ne0 bool convention (chantier 5 review minor).
61 tests: all VizConfig defaults, four bool conventions,
numeric cast fallback, monkeypatch round-trip, and Part B
gating (_ensure_model never called under --iphone-usb).
Fix hand_swap_lr from_env default: was True, now False.
Add data_only_viz/config.py: frozen dataclass VizConfig with typed
fields for all 56 env reads in data_only_viz. from_env() reads
os.environ at call time — no module-level singleton, test-compat.
Migrate action_head_pub._init_finger_piano and renderer.initWithState_
to consume VizConfig. Keep renderer.py ARKIT_FULL module-level (test
uses MetalRenderer.__new__ without calling initWithState_).
HAND_SWAP_LR default changed False (was True): Vision chirality
validated correct live 2026-07-02; swap knob now defaults off.
Review C1 (Critical): /pose/hands per-voice L/R mods routed without
the HAND_SWAP_LR swap while pinch/strike used it — the two paths
contradicted each other on an inverted-chirality source. step() now
takes swap, the publisher passes its env-read value, two tests pin
the contract. Also hoists the route_hands imports (M5) and fixes the
stale slot comment (M1).
New module hand_slots.route_hands():
- Chirality path: iPhone ARBodyTracker Vision 0/1 -> L/R slots,
alignment preserved before invalid-hand filtering.
- Fallback: ascending cx sort (mirror-aware).
- swap=True: swaps the two slots at the very end (safety knob,
HAND_SWAP_LR=1, if a source inverts Vision chirality).
- near_min gate: drops slots below wrist->middle-MCP distance.
18 new tests covering all branches.
Extract apply_skeleton_joints() as a pure module-level function that
copies prev_arr (or zeros) and writes a new array, never mutating the
old one. The TAG_SKELETON handler now stores the new array under the
state lock, so readers that captured a reference to the previous array
(multi.py shallow-copy + ArkitFuse.apply outside the lock) always see
a complete, consistent snapshot — eliminating the B4 tearing race.
6 new tests cover: shape, identity (is not), value preservation for
invalid joints, value application for valid joints, wrong-shape reset,
and no-mutation of prev_arr.
Review follow-up: make the v2 filter's rejection of legacy 0.707
diagonal fixtures explicit instead of silently retuned, and document
the close-up caveat on max_len (ARKIT_BONE_MAX raises it live).
HandPersistenceGate (min_frames=3, radius=0.15) suppresses ghost
hands that flash for fewer than min_frames consecutive frames.
Gate is called once per frame in the renderer on persons_hands;
result filters both the video overlay loop and the side panels.
hand_plausible now rejects wrists outside [-0.1, 1.1] (out-of-frame
ghost anchor check). panel_segments gains a conf_min param.
Env knobs read at renderer init (single parse, no per-frame cost):
HAND_CONF_MIN (default 0.45, raised from 0.3)
HAND_PERSIST_FRAMES (default 3)
ARKIT_BONE_MAX (default 0.5, passed to arkit_segments)
New tests: 12 gate tests + 4 wrist-bounds tests.
arkit_segments now accepts max_len and bounds params.
Bones whose endpoints fall outside [-bounds, 1+bounds] or whose
length exceeds max_len are skipped (tracking-loss garbage).
Defaults: max_len=0.5, bounds=0.1. Renderer will pass
max_len=ARKIT_BONE_MAX explicitly (Part 3).
Tests updated to use realistic bone lengths; new tests added
for the out-of-bounds and overlong-bone cases.
NSEvent keyboard monitors were placed at the END of
_start_pose_worker, unreachable: every backend path returns
early (lines 369/384/399/413/440) so monitors were never
installed. ESC, space (concertNext) and keymaps were silently
dead in all real concert paths.
Move the local + global monitor installation into
applicationDidFinishLaunching_, after window focus setup,
non-headless path only. The --multi-hmr headless early-return
(line 95) is unaffected.
_read_sources used .get("pid")/.get("v3d") (dict API) but
multi_hmr_worker emits SMPLXPerson dataclasses (field: vertices_3d).
AttributeError was swallowed by LOG.exception at 30 Hz.
- Dual-read via isinstance check: dataclass path uses p.pid /
p.vertices_3d / p.expression; dict path keeps p.get() unchanged
- 3 new tests: real SMPLXPerson emits OSC, expression field accepted,
legacy dict format still works
persons_arkit_2d_t was written by IphoneUSBSource but never
read. A USB drop left the frozen skeleton on screen indefinitely.
- Add arkit_2d_fresh(ts_by_pid, now, max_age=1.0) pure helper
in hand_display.py (8 TDD tests, all green)
- renderer._update_skeleton: use_arkit now requires at least one
pid younger than 1 s (replaces bool check)
- Per-pid freshness gate in draw loop skips stale pids
Add panel_rect/panel_segments/panel_frame helpers to hand_display.py
(PANEL_SIDE=0.30, PANEL_MARGIN=0.02, PANEL_INNER=0.10). Segments use
uniform pixel-space scale so the hand aspect ratio is preserved
regardless of view aspect. mirror=True flips X within the panel.
In renderer._update_skeleton add push_panel (no mirror flip) and a
panels block before the segs==0 epilogue: routes persons_hands to
left/right panel via chirality list when aligned, screen-cx fallback
otherwise. Draws frame (pid 7) + wireframe (pid 5 left / 6 right).
Runs on every path including use_arkit. Guards: _mp_bones None, empty
hands, misaligned chirality, height==0.
26 tests pass (10 new for panel helpers).
The AV-Live data-only text HUD hid the lower-left corner of the
performance display and would collide with the new left-hand panel.
VIZ_HUD=1 re-enables it; the H key toggle guards the absent view.
_decode_hands now returns (hands, chirality) tuple instead of a bare
list. Landmark .z keeps the raw wire value; .c = clamp(z, 0, 1).
Chirality byte (1=right, else 0=left) is decoded per hand and stored
in State.persons_hands_chirality under the same lock as
persons_hands_iphone.
L'électron rare