Sets — locations × seasons

What month= actually does. It sets H₀ to that month's p75 significant wave height at CDIP SC116 (MOP v1.1, whole years 2000–2024, 218,975 hours used), de-shoaled from the 15 m contour to the deep-water H₀ the shader re-shoals from. Size only. Period is seasonless in this data — the interpolated spectral peak is 14.4–15.2 s in every month — so a month restores the site card's own T, chop and Δf and changes nothing but height. Tide is not a CDIP product and is left alone. month= and day= are mutually exclusive; an explicit h0= beats both. August is the flat one for a measured reason: across 2000–2024 there are zero hours at or above Hs 1.3 m in July or August. See docs/research/PP_CDIP_CLIMATOLOGY.md.

The requested and drawn heights are not always the same. Monthly p75 is routed through the shipped peel floor before rendering. In this sampled matrix Sewers raises January, October and August to the same 1.54 m drawn H₀; Second Peak keeps January at 1.245 m and raises October/August to 0.47 m. Identical Sewers season rows are therefore a visible cost of the current model policy, not stale captures. Every row header prints both heights — asked (the climatology) and drawn (what the frames show) — with the verdict beside them: floor applied where the asked height reaches the reef but not a peel, or reef inactive … n/a as a season where it does not reach the reef at all. That second case is the summer at Sewers and Second Peak: the reef's own activation height (the depth of the wedge's shallowest cell — Sewers 1.24 m, Second Peak 1.00 m at tide 0) sits above the August p75, so the season there is a lull and the frames in that row are the floor's wave, not August's. The requested climatology also remains in the row label and permalink. See docs/MODEL.md §4.6, docs/research/REEF_ACTIVATION_2026-09-01.md and the clamp control.

What is on this sheet. The location axis is all seven presets at month=january (7 rows). The season axis is two presets × three months (6 rows): January the peak month, October the autumn shoulder, August the flat one. privates is the down-point site: mapped on 2026-09-02 (truncated contour, 1.87 m RMS), but its synthetic reef fit does not converge (7.6° against 31°) and its wedge activates at 0.72 m, above the 0.70 m card, so the card-state line is the surveyed platform's own peel and the app's readout says reef synthetic. Camera is cam=drone throughout: a set is a property of the whole lineup, and the overhead view is the one that shows the envelope arriving down the point.

What was left out. 7 presets × 12 months × 5 clocks would be 420 frames. The sampled 7 + 6 = 13 rows cover the location axis completely at one month and the season axis at its two extremes plus a shoulder, on the two sites with the most contrast in ξ (Sewers 1.15, Second Peak 0.65). Not sampled: the other five presets across months; the nine unsampled months; any tide or Δf variation (a month touches neither).

Why the crest number is swept, not sampled (2026-08-19). A set sheet is about the envelope, and this sheet's columns are one quarter of a set beat apart — 31–42 s — against a carrier wave of 12–15 s. Those two periods have no relation, so a single-instant crest at one station lands on an arbitrary point of the passing wave in every column. It looked exactly like a model defect: The Hook read 1.04 / 3.63 / 3.13 / 1.25 / 0.98 m across the beat, a set apparently peaking a column early and half over by the peak column. Sweeping that same station every 1 s across the beat found waves of 2.11–5.30 m with the biggest arriving at t = 184 s, 3.9 s (0.027 beat) before the peak column's own clock — the envelope was where it was designed to be and the instrument was reading between two waves. The headline crest±T/2 is now the max over one carrier period centred on the column clock; the instant is printed beside it because that is what the frame shows. Nothing in the model moved. See docs/research/MEASUREMENT_LESSONS.md 12.

Why the lull is not flat. #env floors the set envelope at 0.15 instead of exactly zero — env = (1−m) + m·cos(…), m = 0.425, so the peak is unchanged at 1.0 by construction and only the trough rises. That floor is derived from the SC116 spectra two independent ways, not picked. Columns 1 and 5 are the lull; if they read as dead flat water, that is the regression this sheet is for.

Where this sheet says n/a, and why. Two measures go blank at any site with no baked break line. foam_pix needs a baked line to project a corridor onto; with no measured bed there is no baked line, so there is nothing to sample and the cell reads n/a instead of sampling an arbitrary band of pixels. ceilM, the depth-limited crest ceiling in the JSON, is null for the same root cause one step further on: such a site runs u_depthMix = 0, the seabed sampler is a 1×1 stand-in, and the depth the shader reads back is the storage format's quantization floor rather than a seabed — so γh never binds and the "ceiling" would be 1.878·H₀ wearing a depth limit's name. A site with no measured bed has no ceiling to be over. Privates was that site until 2026-09-02, when its bed was mapped on a truncated contour window (1.87 m RMS); an earlier sheet did print the degenerate number there and produced a headline defect ("2.2× over its ceiling") that was entirely an artifact of dividing by it. On a build from the mapped bed every Privates cell carries real reads — its ceiling at the card is 1.41–1.58 m of water under the line and its set-peak pocket fill is 1.17, at the top of the mapped family — and n/a should appear nowhere on these pages; if it does, a preset has failed closed and this note explains the blank.

Method caveat. Column 1 of every row is captured from a cold load at its own #sim=; columns 2–5 advance the clock with __pointbreak.setSim() plus two rAF ticks — the mode scripts/capture_temporal.mjs validated against per-frame reloads. Each cell's hash reproduces its frame either way.

Camera drift, recorded not assumed. The #aim cameras frame the baked line's action centroid and smooth over ~6 s of sim time, so advancing the clock could move the instrument between columns of one row (MEASUREMENT_LESSONS 11: an instrument that frames itself on the signal is not a fixed instrument). Each row header carries the largest camera displacement between any two of its five columns. A row with a non-trivial drift there is comparing slightly different windows, and its numbers should be read accordingly. The camera legitimately does differ between rows — the break line moves with H₀ — so only within-row drift is a concern.

Stills, not motion. Nothing here is a motion read off the pictures (MEASUREMENT_LESSONS 1). The positions in the cells and the peel in the row headers are model reads at clocks the model itself derived, and the frames are ordered by that clock, so the ordering is established rather than inferred — which is precisely what lesson 1's third failure ("peel direction is rightward, confirmed across two frames") lacked. Read as pictures, these frames still cannot tell you which way the peel runs.

← all QA sheets in this build · raw measurements (JSON) — the same provenance block, machine-readable · the essay · the repo · commit b78be48