Add figures/episode_previews/index.md — collapsible per-episode preview picker.

Add 2026-05-19 episode previews; drop stale 2026-03-23 + 2026-05-15 previews (excluded sessions).

Wave 2: multi-cam migration complete for 2026-05-10 + 2026-05-11. All 32 episodes now ship under episodes/ and segments/ with view_left / view_middle / view_right + tactile_{left,right}. Legacy processed/mode1_v1/ and processed/mode2_v1/ removed.

Publish 2026-05-19 session + introduce multi-cam schema (view_left/middle/right). 5 episodes + segments; bad_frames + tasks + README + schema doc updated. Multi-cam rollout for 2026-05-10/11 still in progress (legacy mode1_v1/mode2_v1 paths remain valid until that wave finishes).

examples/play_react_pt.py: accept episode-stem paths. You can now pass .../2026-05-11/episode_005 (no .pt suffix) and the player resolves all matching segments, loads the parent date folder so N/P navigation still works, and seeks to that episode on start.

docs/curation_pipeline.md: point at MultimodalData/twm/scripts/ now that one-shot scripts have been migrated out of /tmp/ into a versioned home.

examples/play_react_pt.py: episode-aware viewer. Accepts a directory of segments — all same-episode segments concatenate into one playback timeline; N/P switches between episodes. Red border + red trail-plot mark on segment-boundary frames. Also supports --save_video_dir for per-episode batch MP4 export.

examples/play_react_pt.py: interactive .pt viewer (same key bindings as twm.visualize, adapted to the downsampled .pt content). Supports both mode1_v1 episodes and mode2_v1 segments. Self-contained (numpy + torch + cv2). Headless MP4 export via --save_video.

Add mode2_v1 schema — 73 pre-sliced clean segments from 27 source episodes (104.7 min, bad intervals excluded by construction). New ReactSegmentDataset + demo. No bad_frames.json filtering needed at the dataloader; data is constructively clean. mode1_v1 stays for backward compatibility. See segments.json + new README section.

ReactWindowDataset: prominent contact-rich window count + stats attribute (ds.stats) + per-filter rejection breakdown in the summary print. Demo: write H.264 MP4 per picked window (was just static PNG); --no_mp4 to skip; --mp4_fps to tune.

Re-render 15 real_still candidates with the new per-episode p01 GelSight reference (the .gs_ref_{left,right}.png files just shipped). Diff thumbnails now show absolute contact intensity vs the quietest moment of the episode, not change-since-clip-start.

Add per-episode GelSight reference images: episode_NNN.gs_ref_{left,right}.png = the 1st-percentile-intensity frame at native 640×480 (the 'quietest' / least-contacted moment of the episode). Use these as the diff baseline in any visualization — gives a stable 'this is what's currently pressing on the gel' diff rather than 'change since the start of this clip'. 27 episodes × 2 sides = 54 files.

real_still: render ALL 15 candidates with relaxed thresholds (pose<=20mm/s, win=60, view<2.5); 1x playback, 2s window+1s pad each side. candidates.json lists per-event motion stats.

real_still: re-render at 3s window length + 1x (real-time) playback; drop the previous 5x 2s versions. Only 1 candidate meets the strict 3-second still criteria (operator was actively manipulating almost the entire 106 minutes).

Add real_still candidate clips for verification: only 2 windows in the whole dataset meet the strict still criteria (pose<=10mm/s, cam2<2, tac<3) — telling us the operators are actively manipulating almost the entire time. Manifest in candidates.json.

freeze_diagnose threshold: TAU_VIEW_STILL 4.0 -> 2.0 (more aggressive "is the scene really still?" check). Verdicts unchanged for current 41 events because they all trigger ot_loss via other signals.

Drop 11 stale ot_loss MP4s from the pre-opt_gap-fix run (their bad_frames intervals are still valid but the verdict-label numerics in the filename matched the buggy 47-event sampling, not the corrected 41-event sampling).

freeze_diagnose: fix max_opt_gap() returning 0 when OT stream silent; drop events entirely inside trimmed prefix; remove misclassified real_still/ (those were ep_005/017 OT-offline prefixes that shouldnt have appeared at all). Final: 41 events, all ot_loss.

Repopulate freeze_diagnose/ MP4 clips (folder had become empty after the GIF-purge): 15 ot_loss + 4 real_still samples, 5x speed yuv444p MP4. Refresh summary.json (new run: 43 ot_loss + 4 real_still events; previously labelled all 47 as ot_loss).

sample_grid.png re-rendered at window_length=32 (was 16) + 4 row-matched MP4 clips (sample_grid_clip_NN.mp4) so each grid row has a video twin showing real recording motion at 5x.

ReactWindowDataset: default min_motion_fraction 0.5 -> 0.25 (a window passes if >=25 % of frames per active sensor are above the speed threshold).

ReactWindowDataset: default min_motion_mps 0.03 -> 0.01 (30 mm/s -> 10 mm/s). Recipes + README + demo + CLI updated.

MP4s re-encoded at 5× speed (yuv444p / 8-bit / CRF 20 — sharper text overlays); all 66 GIFs removed from HF (MP4-only going forward). Source 8-bit RGB → 10-bit would be wasted bits; 4:4:4 chroma is the upgrade that actually matters for the text panels.

README cleanup: fix stale 75 % bimanual-contact claim (actual 64.3 % post-trim); reference freeze_intervals.json + MP4 previews + metadata/episodes.parquet that previous batches added.

Re-render 20 sample_window GIFs + MP4s with the trim_offset bug fixed (was showing OT-offline prefix region for windows from trimmed eps; now shows the correct post-trim region with live OT pose updating across frames).

Audit batch 2: regen contact_intensity hero + F1/F2/F8 + freeze_diagnose summary with post-trim numbers (#1, #2, #6). 3302 contact events post-trim (was 4136 pre-trim).

Audit batch 3: ship .mp4 alongside every GIF for inline HF rendering (#11). 51 files, total GIF→MP4: 451→24 MB.

Audit batch 4: per-episode metadata Parquet at `metadata/episodes.parquet` + HF data-files config so the dataset-card viewer renders a sortable table (#13).

Audit batch 1: post-trim numbers in docs/figures.md (#3); tasks.json 2026-05-11 note + trim_offsets lookup (#5, #8); bad_frames.json trim_offsets header (#8); dataloader default which_sensors='both' (#10); new freeze_intervals.json with per-event diagnostic signals (#12)

examples/: upgraded ReactWindowDataset (adds motion filter to catch operator-paused windows the bad_frames check cant); canonical self-contained demo_react_window.py (no twm dependency); new examples/README.md with four-filter explanation and recipes per use case

Update README + docs/{quality,caveats}.md: post-trim totals, three concrete filtering recipes (example dataloader / direct .pt / pose-only), rewritten OT-track-loss section reflecting the prefix-trim + recorder watchdog fix

Add 20 dataloader sample-window GIFs (random windows from post-trim ReactWindowDataset) for visual ot_loss verification

Re-render previews for 3 trimmed episodes (ep_005/012/017) starting from post-trim region

Trim OT-uninitialized prefixes from 3 .pt files (ep_005/012/017 — see `_contact_meta.trim_offset`); recompute bad_frames + data_quality_breakdown in trimmed coords; ot_loss frame count now reflects real mid-episode mocap dropout only. Recorder-side OT watchdog added to data_collection.py (in twm repo) so future episodes won't repeat the bug.

Bake ot_loss intervals into bad_frames.json (47 events, rotation-aware detector); dataloader skips them; docs/quality + caveats updated; data_quality_breakdown #3 renamed sensor_at_rest → ot_loss

Re-label freeze classifier: events are OptiTrack track losses (ot_loss), not recorder-pipeline stalls. GelSight stream is healthy throughout; tac_dup was misread (gel only changes under contact deformation).

Visualization refactor (twm/viz.py single source); native 1280x480 + left|middle|right cam order; freeze_diagnose/ replaces freeze_check/ and freeze_classify/; finding: all freezes are recording-pipeline stalls

Add freeze_classify/ — cross-modal motion classifier for sub-second freezes (15 samples per class)

Re-render freeze_check/ in recording-viewer layout (3 cams + GS thumbs + projection); add 1 s pre/post pad; onset/offset/mid scheduling for long freezes

Add freeze_check/ — 55 real-time OptiTrack-freeze inspection clips

docs/curation_pipeline.md: add BGR-vs-RGB channel-order gotcha

Drop 2026-03-23 right-only pilot (outdated); recompute totals across 27 bimanual recordings

Fix RealSense BGR→RGB; sample_window.gif now matches recording-viewer layout

Regen previews at 256-color palette; 3-cam dataloader gif; add docs/curation_pipeline.md

Add example dataloader: ReactWindowDataset for short contact-rich window sampling + demo (static grid + GIF) with sensor-frame overlay

Lead with robot-arm-free / human-hand collection — README hero + 'What's different' section, tasks.json gets operator/embodiment fields, recording.md explains setup, F7 adds Operator column

Reframe dataset as dynamics / world-model interaction data (not policy demos); document 2026-03-23 as right-sensor-only pilot; add active_sensors per-date in tasks.json; lead with hours and contact frames instead of episode count

Split README into lean front page + docs/{recording,schema,quality,figures,caveats}.md

README: embed inspection figures inline (intensity overview, teleport reports, data-quality summary)

Add data-quality docs: bad_frames.json, per-mode breakdown, inspection figures; refresh episode GIFs with cam2 projection overlay