The display write dithers, and every baseline moved once

All 27 baselines were re-blessed on 2026-08-12, in one commit that contains nothing else. If you find that commit in the history and wonder what happened: the tonemap now adds ±1 encoded LSB of triangular noise before the 8-bit write (ADR-0096), hashed from the fragment’s integer coordinates and divided by the sRGB transfer function’s local slope. It is always on and it is not a param — correct quantization of the display write is not a look. So every pixel in the engine can shift by one level, and every baseline did.

The re-bless is bounded, and that was asserted rather than trusted. round(x + n) with |n| ≤ 1 differs from round(x) by at most one level, so the whole change is provable rather than eyeballed. Measured bless-to-bless across all 27 — a control set blessed from the pre-dither commit on the same box first, because 8 of the 27 rewrite against their committed bytes on a clean local bless and comparing against the repository would have charged that drift to the dither:

channels compared	2 049 408
max |before − after|	2 (WARP) / 1 (hardware)
channels moved	11.52 %
channels moved by 2	0.0103 %

The two-level moves are the blessing adapter’s, not the amplitude’s, and this is worth knowing before it is rediscovered. On the hardware adapter the bound is exactly one: zero of 12 288 channels move by 2 on flat-sweep probes at either end of the range. On WARP, 212 channels across the 27 baselines move by 2, 88 % of them below byte 20 and every one skipping exactly one value. WARP is not missing those code values — an undithered ramp there contains every byte from 6 to 18 with no gaps. DX12 permits tolerance in float-to-sRGB8, and in the steep dark region WARP’s approximation departs from the true transfer function, so a perturbation sized by the true slope lands two levels away in some places and fails to move the value at all in others. Below ~byte 20 a WARP capture is not a reliable instrument for one-level effects; take those on hardware.

The guards live in core/src/render/tonemap/tests.rs: the_dither_is_one_encoded_level_at_both_ends_of_the_range (the amplitude, which is what a “tidied away” slope term breaks) and the_dither_dissolves_a_dark_ramps_plateaus (the banding itself, stated as a ratio against an undithered control resolved in the same run).

Two pixels per 8-bit level is the SAFE case, not the dangerous one. Plan 0080 Phase 7 reasoned the other way — it called a quarter-frame fade at roughly two pixels per level “the classic Mach-band configuration” — and the arithmetic is inverted. Banding lives where one level lasts a long time, which is the flattest part of a curve, so a dense packing is the healthy state and a bg_ramp_gamma below 1 (a long dim tail) is where to look. That plan is closed and its own text is history; the correction belongs here, where someone reading about plateau widths will meet it. The reference frame and its before/after are in core/tests/fixtures/scratch-0082/.