lp-ws281x: compute the RMT block plan at driver init from the manifest's declared channel count

[Yona-Appletree]

What

Replaces the compile-time BLOCKS_PER_CHANNEL constant in the lp-ws281x chip
backends with a block plan computed once at driver init from the board
manifest's declared /rmt/ws281xK count. One build now serves every board
shape; the ws281x_2blocks cargo feature is deleted.

Rule: N declared channels on a chip with B usable blocks → each channel gets
floor(B/N) consecutive blocks; absorbed slots are never configured (same
construction as before); remainder blocks stay idle.

Chip	Declared (shipped manifest)	Plan	Window / half	vs. before
C6 (XIAO)	2	[1,1,0,0]	48 w / 24 w	unchanged (G1 default)
C6 (DevKitC)	1	[4,0,0,0] — RX blocks absorbed	192 w / 96 w (~120 µs)	was 48 w; restores legacy-driver margin
classic (DOM-Z-102)	4	[2,0,2,0,2,0,2,0]	128 w / 64 w	byte-identical plan; P4 baseline comparability preserved
classic	1	[4,0,…] (cap: tx_lim is 9-bit)	256 w / 128 w	new
S3 (XIAO Plus)	4	[1,1,1,1]	48 w / 24 w	unchanged
S3 (DevKitC)	1	[4,0,0,0]	192 w / 96 w	was 48 w

Why

RMT refill misses tear frames, and margin (bigger halves) is the only software
lever left at Priority::max() (RMT-priority plan F1). Measured on the C6:
28 % of frames truncate under WiFi scan at 24-word halves vs 0.49 % at 48
(docs/debt/c6-scan-truncation-accepted.md). A single-strip board was paying
those deadlines while blocks idled; the deleted legacy C6 driver gave its lone
channel all four blocks. Since the manifest is already the sole authority on
channel count (ADR 2026-07-31), the split follows from it — per Yona: "if
users are using only 1 strip, we should be using the whole buffer."

How

• lp-ws281x core: BlockPlan::for_channels(n, width) (+ nth_available
  / index_of_slot slot mapping) and SharedBlockPlan — an atomic,
  set-once-at-init, fail-closed slot each chip backend reads (windows can
  never change after init, honoring the RmtHw::ram_words stability
  contract). The driver core drops its duplicate plan copy:
  ram_words(ch) == 0 now means unavailable, making the backend the single
  source of truth (with_blocks/block_plan() removed).
• C6: plan ranges over all 4 blocks; a lone TX channel absorbs the RX
  pair — verified against esp-hal 1.1.1 (reserve_channel_memory bounds
  memsize by NUM_CHANNELS - channel = 4 for ch0, absorbed channels marked
  Reserved) and by the legacy driver's years on this silicon. Harness
  LedChannel publishes the 1-channel plan, so every test_rmt-class run
  exercises the widened window on silicon.
• classic: per-channel cap of 4 blocks — tx_lim is 9 bits, a 512-word
  window cannot be expressed.
• S3: gains the classic/C6 manifest-index → RMT-slot indirection (needed
  once plans can widen); loopback harness publishes the 4-channel plan
  explicitly so its on-silicon expectations (TRUNC_EXPECT_BITS etc.) are
  unchanged. RX blocks left alone (loopback uses them).
• Telemetry line format unchanged — half= simply reflects the computed plan.

Validation

• cargo test -p lp-ws281x — all suites green, including new host tests for
  the plan arithmetic (shipped shapes, windows/stride, absorbed slots, odd
  remainder, SharedBlockPlan set-once/fail-closed).
• Clippy -D warnings: fw-esp32c6 (+ all 14 harnesses + ws281x_telemetry),
  fw-esp32v3 (+ ws281x_telemetry), fw-esp32s3 (+ test_loopback).
• Size checks: c6 2,881,856 B (263,872 B headroom), v3 1,727,840 B, s3 PASS.
• just fw-manifest-check-esp32c6 — embedded manifest unchanged.
• just check test — PASS.

Hardware

Desk boards were not connected during this session (no usbmodem* /
wchusbserial* devices present) — hardware smoke is pending: C6 jig 2-channel
default + a temporary 1-channel manifest for the 192-word window; classic
quad-strips unchanged-outcome check.

Notes

• docs/debt/c6-scan-truncation-accepted.md lands in docs: RMT-priority plan close-out — G1 decisions, license addendum, C6 scan-tradeoff debt entry #271; its exit-criteria
  update (1-channel manifests now self-serve the wide config; the 2-channel
  tradeoff remains carried) follows here once docs: RMT-priority plan close-out — G1 decisions, license addendum, C6 scan-tradeoff debt entry #271 merges, to avoid an
  add/add conflict.

🤖 Generated with Claude Code