The notebook becomes a course, and finds the bug its zero was hiding

The 259-cell encyclopedia is now a 28-cell introduction that starts
with a video in your hands and ends with sound against motion: load and
look, the average image and the two grams, the QoM curve, the body, the
segments, the audio, the audio-motion tempo comparison, and where to go
next --- executed once, images only, 6.8 MB where 16 stood. Building it
exposed a real fault: pose() caches flat 67-value rows, and both
postures_from_pose() and the actions pose path consumed them as
landmark arrays --- zero postures for every recording, and a hard
failure for actions. pose_cache_landmarks() restores the trajectory
contract with refusals marked invisible, both consumers use it, and the
actions path unpacks pose_qom as what it is: a scalar QoM beside the
cuttable speed envelope. The dancer truly holds two postures, an
opening stillness and a held moment at 44 s, and the notebook now says
so. The changelog merges this session's Fixed entry with the parallel
session's Unreleased additions; that session's code stays its own.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRX1cv8m5PFVZGjfWZUzJj

Author: alexarje

CHANGELOG.md

@@ -7,6 +7,20 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed
+- **The pose cache reaches the higher layers in the right shape** —
+  `pose()` caches one flat row per frame (`[time_ms, x0, y0, ...]`, zeroed
+  below threshold), but `postures_from_pose()` and `actions_from_motion()`'s
+  pose path consumed it as a landmark array: the postures segmenter hit its
+  shape guard and returned zero postures for every recording, and the actions
+  path failed outright. A new `pose_cache_landmarks()` restores the
+  `(frames, 33, 3)` contract with visibility 0 marking the zeroed refusals,
+  both consumers use it, and the actions path also unpacks `pose_qom`
+  correctly (the QoM is a scalar; the cuttable series is the speed envelope)
+  and masks refusals as NaN rather than leaps to the frame corner. Measured
+  on the bundled dancer: two genuine postures (an opening stillness and a
+  held moment at 44 s) and 18 actions where both previously reported nothing.
+
 ### Added
 - **ReID v2.1: chains across breaks** — `associate_fragments(embeddings=...)`
   now links each segment's movers into persistent chains by the same
@@ -22,6 +36,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   segment to a 12.7-minute chain over five segments, with 119 chains spanning
   breaks. Zero new decodes, as priced in the v2 design.
 
+- **Events against events** — new `_events` module. `event_alignment` measures how far each
+  event of one stream (strokes, footfalls, looks) falls from the nearest event of another
+  (note onsets, beats, cues) against uniformly placed surrogate references, and says whether
+  the two *attract*, *avoid* or sit at chance, with the signed lag of the nearest reference
+  telling which came first; `event_xcorr` correlates the binned event trains. On the
+  painter–pianist session this separated three takes no envelope correlation had.
+- **Cross-recurrence** — `_correlate.cross_recurrence`: recurrence rate at a fixed radius
+  quantile, determinism, mean line length and the diagonal recurrence profile of two series,
+  each against circular-shift surrogates, so a high determinism from two smooth series is not
+  mistaken for coordination.
+- **The painting as a time series** — new `_canvas` module. `painting_content` reduces a
+  canvas video to one hand-free median frame per second and measures painted share, monotone
+  coverage, chromatic share, the hue histogram (drawn as a *colourgram* by `colourgram_image`),
+  warm/cool balance, edge density, composition (`composition`: paint centre and spread,
+  left–right symmetry, edge-orientation histogram and anisotropy) and a dominant-colour
+  palette per minute. Bound on `MgVideo` as `painting()`.
+- **Head turns from the IMU** — `_pupillabs.head_turns`: yaw relative to a running median,
+  and the spans where the head is turned away from what the wearer has been facing.
+
 ## [1.31.0] — 2026-09-03
 
 ### Added