ADR-008 generalized: flags bits 6-7 = (flags >> 6) & 0b11 — 0 none /
1 G (v1.11's bit-6 shape, every old CD file decodes unchanged) /
2 R / 3 B; FLAG_REF_R = 128; reserved mask widens to bits 8..31.

channel_diff_relabel/unrelabel generalize per ref_pick (planar
per-plane in wire emission order; token tag-0 + interleaved mode-0
per pixel); the encoder pick sweeps reference 0..=3 across every
record arm and the token arm at complete-file cost.

Gate: corpus-gate 832,765 (−2,120), release-CLI e2e 30/30 verify
832,739 (−2,146 = the C2 census prediction reproduced within 2%);
REF dist none 7 / G 13 / R 5 / B 5; encode ~1.5x, decode unchanged;
compatibility matrix verified against the released 1.11.0 binary
(REF=0/1 exact both; REF=2/3 v1.11 reject / v1.12 exact); conformance
10/10 (REF2/REF3 goldens + the bit-8 reject boundary replacing the
bit-7 vector the mask widening invalidated); fuzz 1,025 trials
0 panics (4 accepted = diffed benign valid-different-file class);
make gate OK.

Census caveat recorded in the ADR: the C2 per-file repack used
python-zlib-9 against the shipped flate2-default(6) container —
level-mismatched per-file deltas (document's −166 = phantom), the
aggregate prediction survived.

Jon <jon@nocturnalinc.com>