SSD1333 brightness: dimming policy refit from the measured response surface (stage 3)

[brickbots]

Stage 3 of the panel-characterization effort (follows #568): DisplaySSD1333.set_brightness is refit against the rig-measured response surface, and the shipped curve is verified end-to-end on the photometer rig.

What changed

The old policy factored a target into per-axis multipliers — a model the stage-2 photometry falsified (the axes don't multiply; contrast and master are one drive-product axis; pre-charge authority runs 32× to 1.1× depending on drive). The new policy walks three measured response tables, one per regime:

levels	regime	axes	measured span
1–20	pre-charge	ceiling 4, drive at cut-out, pc code = level+3 (4→0x17)	151 → 5.03e4 ADU/s
21–47	ceiling	drive at cut-out, pc full, ceiling = level−16 (5→31)	6.43e4 → 2.03e5
48–255	drive	full ceiling/pc, power-law target inverted through measured f(P)	2.03e5 → 1.05e6

Tables come from a dedicated same-day calibration session (stage3-calibration-20260802.jsonl, 74 points; repeat-point agreement 0.02%, reference within 0.9% of stage 2). Standing user decisions kept: dim-weighted knee shape, tonal-range rule (ceiling ≥ 4, value-64 canary), MAX_CONTRAST = 160 and the 70% blooming cap (eyeball constants, unexamined), top at ~75% of clean max, panel-agnostic harness.

Rig verification (acceptance test)

Selftest gates all PASS, then one measurement per level, all 255 levels, axes exactly as the policy programs them (stage3-verification-20260802.jsonl):

• Bottom = the panel's floor: level 1 measured 149.6 ADU/s (measured emission floor 148–151). The old policy's dimmest state sat 40× above this.
• Top: level 255 measured 1.046e6 ADU/s, +0.3% of the soft-top target.
• Span: 6998:1 — 3.84 decades within the tonal rule (old policy: 233:1).
• No dead keypresses: under the UI's real +20%/−10% stepping, every press changes the light, both directions, all levels.
• Four adjacent-level dips of ~2.5% at levels 76–112 are rig noise, not policy inversions: those fluxes saturate the period-locked tier and fall to the sub-period tier, whose 48-frame mean carries ~4.6% beat noise (single-frame CV ~33% in the journal). The programmed states come from monotone measured tables. Sentinels show only a slow +2.6% thermal warm-up across the ~35-minute run.
• Model vs measured: median 0.75% across the curve; worst ~7%, confined to the sub-period tier + warm-up region.

Saved settings shift (like the 2026-07 reshape — call-out)

The knee state is unchanged, so settings ≥ 20 render within ~30% of before (old 50 ≈ new 48, old 125 ≈ new 96; the top is 18% dimmer under the soft cap). Below the knee the range runs 40× deeper: the old dimmest setting's light now lives around level 8, and levels 1–7 are darkness the old model couldn't reach. Field users who lived at the old bottom will want to press up a few times.

Docs

• ADR 0023 amended in place: separability, pre-charge linearity/cut-out/"dark at 0", the 0.16% drive-floor claim, and the 6%/35% curve description all replaced with measured facts; decisions unchanged.
• Display CONTEXT.md: knee-curve entry refit, ceiling/drive regime terms added.
• ssd1333-response.md §7 marked resolved; stage-3 handoff file retired (fulfilled).
• Unit tests rewritten around the measured tables and real keypress trajectories (17 pass; full unit suite 1151 pass).

🤖 Generated with Claude Code