Session B: local rain checkpoint

2026-10-18 15:36 start: bitcube · Opus 5.5

Full plan: B-plan.md.

Steps:

• B1 Structured-light capture: Gray-code gray/white sources in artnet_out.py; tools/pixelmap/capture.py (one ffmpeg reader, two passes, retakes on disagreement, brightness 0.4).
• B2 Decoder (TDD, pytest, .venv): camera pixel → (rope, px, conf); per (rope, px) median (u, v); outlier filter; fill along the rope; tools/pixelmap/pixelmap-2026-10-18.json and overlay.
• B3 HUB75 quad 1 corners in camera space from the panels test pattern.
• B4 Rain/river field (built-in TOPs only, 1280×720): noise river with drift and breathing, plus feedback rain streaks.
• B5 Sampling: UV table → TOP to CHOP → 20×400 rope_canvas → ropes Canvas; Corner Pin from the B3 corners → panels Quad1.
• B6 river-local-oct-18.toe (Save As of river.toe), td/build/build_river_local.py, td_save.ps1 -Target (default = the open file). Autostart unchanged.
• B7 Verify by camera (stills and a clip), tune, report watts and fps.
• B8 Checkpoint still on the status page, tick L, log, code review, summary.

Progress:

• 15:38 B6 (early) TD saved as river-local-oct-18.toe via the new td_save.ps1 -Target before any building, so river.toe stays untouched. TD now runs the checkpoint file.
• 15:46 B1 ✅ Capture frames/pixelmap-20261018-154155 (29 patterns × 2 passes, settle 1.75 s, drift 13.3, no retakes needed, all pairs agree to within 1.6%). Note: this run still had the all-ropes white reference (≈1.2 kW, both boxes at the limiter, about 10 s in total) because session A’s review arrived after it started. White is now removed from the code; the lit level comes from the bit/inverse pairs. (commit bc4e2f0)
• 15:46 B2 ✅ Decoder (15 pytest tests) → tools/pixelmap/pixelmap-2026-10-18.json: 164k camera px decoded; all 18 working ropes located, 53–214 of 400 px visible each, the rest interpolated along the rope; ropes 9 and 17 absent. (commit 043a8cf)
• 15:48 B3 ✅ HUB75 quad 1 corners found from the panel test pattern vs black (frames/hub75-20261018-154718/corners.png), red at TL as expected, checked by eye. Stored in the pixel map as hub75_quad1. The coils hide part of the bottom edge. (commit 1ebe7d9)
• 15:51 B2 (revised, per session A) Tighter reject_outliers: drops points with no near neighbour along the rope and islands shorter than 8 px between jumps of more than 0.08. fill_along_rope no longer draws chords across a jump between piles; each half of a hidden stretch holds its nearer visible point. The jump to the camera’s status light is gone. 22 tests. Pixel map regenerated from the same capture, with no re-capture and no white frame. (commit 5cd1ac6)
• 15:56 B4 ✅ Rain/river field /project1/river_local (1280×720, camera space, built-in TOPs only):
  • River: simplex noise stretched along the flow, drifting slowly, breathing (14 s) through a black → deep blue → teal → white ramp, more than half black.
  • Rain: thin streak noise scrolling down (0.22 heights/s), thresholded sparse, with feedback trails. The first version (per-frame random seeds) gave stationary dots, so it was replaced.
  • Look parameters live on the COMP (Gain, Riverlevel, Flowspeed, Breathe*, Rain*, Blur, Uvtest). (commit 2310582)
• 15:56 B5 ✅ Sampling: UV table (8,000 rows, stored in the .toe) → TOP to CHOP → CHOP to TOP rowswrapped 400×20 → Flip (flop bottomleft + flipx) → 20×400 rope_canvas → ropes Canvas.
  • Checked with the UV test field: max error 0.0006, and unmapped ropes are exactly black.
  • HUB75: Corner Pin (perspective) from the B3 corners → 256×128 → panels Quad1; all four corners match. (commit 2310582)
• 15:56 B6 ✅ river-local-oct-18.toe saved via td_save.ps1 and committed. Idempotent td/build/build_river_local.py, check script td/scripts/check_river_local.py, Web Render switched off in this file. Autostart unchanged (still river.toe). (commit 2310582)
• 15:57 B7 ✅ Camera verification (clips in frames/, not committed: people):
  • The river washes drift across the left piles, and the rain shows as columns of white dashes falling through successive rings of the centre and right piles.
  • The HUB75 shows the same streaks descending in step with the piles beside it.
  • Tuning: rain density 0.06 and wider streaks (sx 4.5); river inlow 0.5; ropes Brightness 0.4 (cap); panels Brightness 0.5 (0.8 overexposed next to the ropes).
  • Art-Net 68–196 W total (10 samples, both boxes), no dropped frames, about 35 of 40 frames/s sent. TD 60 fps (measured over 10 s).
  • (B5 line fixed: an earlier push had control characters where backticks were.)
• 15:58 B8 ✅ Checkpoint still docs/img/rain-checkpoint-2026-10-18.jpg (cropped, no people) on the status page; plan item L ticked; log entry; CLAUDE.md now lists both show files. (commit 4c3e7c8)
• 16:05 Code review ✅ 10 findings, all fixed. The important one: plain Gray codes are unbalanced, so the 15:42 capture lit b04n (ropes 0–15 fully, about 960 W) and b13n (64% of every rope) at 0.4, twice each. My CLAUDE.md line “patterns light at most half the pixels” was wrong. capture.py now dims each bit/inverse pair so no box draws more than half its power at the cap (at most 300 W/box). Also fixed: Graybit clamp, build guard against river.toe, settings restored after a capture (checked on the hardware), newest-pixel-map rule, td_save with an incremental file open, clear errors, and a TD↔decoder sync test (26 tests). (commit dbd50f9: the fixes landed in the commit titled “B8 done, code review line, SESSION SUMMARY” because a commit message with quotes failed in PowerShell)

SESSION B SUMMARY — Sun 18 Oct 2026, 16:06

Done:

• B1 Gray-code structured-light capture of all ropes with the gallery cam (tools/pixelmap/capture.py, rope Source gray).
• B2 Decoder (TDD, 26 tests) and tools/pixelmap/pixelmap-2026-10-18.json: all 18 working ropes located, hidden pixels filled along the rope, islands and jumps rejected.
• B3 HUB75 quad 1 corners in camera space (hub75.py), red at TL, checked by eye.
• B4 Rain/river field /project1/river_local, built-in TOPs only: drifting, breathing blue/teal/white river plus falling streak rain, on black.
• B5 Sampling: TOP to CHOP at 8,000 UVs → 20×400 → ropes; Corner Pin → 256×128 → panels quad 1. Both checked numerically (error < 0.001).
• B6 river-local-oct-18.toe (Save As of river.toe; river.toe untouched); idempotent td/build/build_river_local.py; td_save.ps1 -Target.
• B7 Camera check: drops fall through every pile as columns of lit rings and the same streaks cross the HUB75 in step; river washes drift across the piles. 68–196 W total, TD 60 fps, Art-Net about 35 fps sent, no drops.
• B8 Still on the status page, plan item L ticked, log and CLAUDE.md updated, code review done and fixed. Not done / blocked: none. Ropes 9 and 17 still dark (P1 deferred). Evidence: commits d4edc83 (plan), bc4e2f0, 043a8cf, 5cd1ac6, 1ebe7d9, 2310582, 4c3e7c8, dbd50f9 (code review fixes). docs/img/rain-checkpoint-2026-10-18.jpg, tools/pixelmap/pixelmap-2026-10-18.json. Overlay, corners and clips in frames/ (local only; people in them). Hardware state now: TD runs river-local-oct-18.toe. Ropes Source=canvas, Brightness 0.4 (cap), showing the rain. Panels Source=canvas, Brightness 0.5, quad 1 rain, quad 2 blanked. RopeLights stopped. The login autostart still opens river.toe (ropes black, panel test pattern), so a reboot leaves the checkpoint. Decisions or answers needed from Gavin/team:
• Should the login autostart open river-local-oct-18.toe until Dave’s visuals are in? Not changed; your call.
• Awareness: the 15:42 capture briefly lit near the limiters (all-white reference, then b04n/b13n, about 1 kW for a few seconds each). It is fixed in code now. Suggested next step: C1–C3 (session C). C3 must save each show file while it is open (td_save.ps1 saves the open file). Then D, which reuses build_river_local.py’s sampling for Dave’s canvas.