`, AND THAT IS NOT A PREFERENCE.** `a-page-in` holds an identity transform for the life of the page, and a transform makes its element the containing block for fixed descendants — a `.pagebg` inside main would pin to main and scroll with it. It is the first child of `` on all four pages. (`NEXT.md` → Traps, first entry.) ⚠ **`z-index` is deliberately NOT set.** A fixed, untransformed `z-index: auto` box paints in DOM order against `main`, which is transformed and therefore also paints at the 0 level — so "first in the body" IS the stacking rule here. 🚫 A negative z-index would drop it behind the body's own background and the tooth would hide it completely. ⚠ **8% is not a starting point to tune by eye at full strength.** An ink wash at .08 over cream is a few values of grey; judged in a viewer at 100% every one of these looks far too strong. Judge them on the page. ⚠ **No paper ground is baked into any of the four files** — measured on import, all four corners alpha 0 and zero matte pixels across all ten deliveries. A future re-roll must hold that: a file with its own cream ground paints an opaque rectangle here and kills the tooth underneath. ── the vertical correction, and why it is a transform ────────────────────────────── Fuji and the coast are bottom-anchored, and **both arrived with empty frame below the ink** — the mountain's last row is 996 of 1200 and the coast's is 1002 of 1200. Anchored `bottom`, the motif would float **16.9%** / **16.4%** of the layer's height above the fold. The files are NOT re-cropped for it: they are WebP on alpha and this lab's encoder would have to re-compress them to fix a number CSS can express exactly. `translateY` of the measured fraction is free and reversible. ⚠ Re-measure it if either file is re-rolled. ── what each one is ─────────────────────────────────────────────────── fuji 2000 × 1200, full width, bottom-anchored. The horizon of the whole site. coast 2000 × 1200, the same treatment on Travel — deliberately the same SHAPE as Fuji so the two pages rhyme rather than compete. bough 1400 × 1800, hanging from the TOP-RIGHT corner, capped in px so it stays an ornament on a wide screen instead of growing into a mural. blossom 900 × 900, one flower, bottom-right, the smallest and quietest of the four — Reply is the site's most functional surface and §19's allocation gives it the least. ⚠ The two corner motifs take **.07**, a hair under the pair of washes: they are small and dense, and at .08 the bough read as a smudge in the corner rather than as ink. */ [data-skin="hanakenji"] .pagebg { position: fixed; pointer-events: none; opacity: .08; background-repeat: no-repeat; } [data-skin="hanakenji"] .pagebg--fuji, [data-skin="hanakenji"] .pagebg--coast { inset: auto 0 0 0; width: 100%; aspect-ratio: 2000 / 1200; background-size: 100% auto; background-position: center bottom; } [data-skin="hanakenji"] .pagebg--fuji { background-image: url("../assets/hanakenji/bg-fuji.webp"); transform: translateY(16.9%); } [data-skin="hanakenji"] .pagebg--coast { background-image: url("../assets/hanakenji/bg-coast.webp"); transform: translateY(16.4%); } [data-skin="hanakenji"] .pagebg--bough { inset: 0 0 auto auto; width: min(46vw, 30rem); aspect-ratio: 1400 / 1800; background-image: url("../assets/hanakenji/bg-bough.webp"); background-size: 100% auto; background-position: right top; opacity: .07; } [data-skin="hanakenji"] .pagebg--blossom { inset: auto 0 0 auto; width: min(34vw, 15rem); aspect-ratio: 1 / 1; margin: 0 clamp(1rem, 4vw, 4rem) clamp(4rem, 10vh, 8rem) 0; background-image: url("../assets/hanakenji/bg-blossom.webp"); background-size: 100% auto; background-position: center; opacity: .07; } /* 🚫 The backdrop does not print. It is a page's atmosphere, and a printed page has the paper's. */ @media print { [data-skin="hanakenji"] .pagebg { display: none !important; } } /* ── 39.1 · ⚠ THE BACKDROP MOVED THE GROUND, SO `--faint` IS RE-POINTED ──────────────── §39 shipped without this and the gate caught it — the same fault as §38.2's, one step later in the same round, and the standing decision written for §38.2 (*A veil measured against `--ink` has not been measured*) is exactly what covers it. **A new ground is a re-measurement of every ink standing on it.** Measured by compositing each of the four WebPs at its own layer opacity over `--paper` and taking the DARKEST resulting ground, then reading the resolved oklab inks against it: ground --ink --muted --faint (62%) bare paper rgb(250,246,239) 13.53 5.89 4.64 bg-fuji @.08 rgb(234,231,224) 11.80 5.13 **4.05** ❌ bg-coast @.08 rgb(230,227,220) 11.37 4.95 **3.90** ❌ bg-bough @.07 rgb(233,229,223) 11.62 5.05 **3.98** ❌ bg-blossom @.07 rgb(233,229,223) 11.62 5.05 **3.98** ❌ `--ink` and `--muted` keep a wide margin. **`--faint` had 0.14 of margin on bare paper and the wash spends 0.59 of it**, and it is live under real small type: `.sched-note` and the `.marg` lines sit directly over where the layer paints. ⚠⚠ **LOWERING THE OPACITY CANNOT FIX THIS, and that is the useful part.** The coast's darkest ink is effectively alpha-1 near-black, so at `opacity: .03` the ground is still rgb(243,239,232) and 62% `--faint` is still **4.36** — by which point the motif is invisible. **The ink is the lever, not the veil.** 🚫 Do not "fix" a backdrop contrast finding by fading the backdrop. So `--faint` goes **62% → 68% sumi**, measured as the floor that clears the worst of the four: coast **4.65** · bough/blossom 4.75 · fuji **4.83** · bare paper **5.54**. (66% was tried first and leaves the coast at 4.38.) ── 39.1b · ⚠ AND `--muted` MOVES WITH IT, BECAUSE THE RAMP IS THE POINT ────────────── The first cut of 39.1 asserted "§A2's ramp survives: 70/68 is still a visible step" and **did not measure it**. It is not a step at all — the gate caught that too, and it is the THIRD unmeasured assertion in this round, which is why the standing decision is worth re-reading rather than re-learning. muted 70% rgb(98, 94,102) oklab L 0.4884 faint 68% rgb(102,98,106) oklab L 0.5026 → ΔL **0.0142**, **1.06:1**, 4/255 a channel. The same grey. §A2's own header names the step as load-bearing — "the ramp survives because 70 / 62 is still a visible step" — and it is live: the ledger's three-tier read (uppercase label `--faint` / value `--ink` / sub-line `--muted`) and eight other adjacent pairs on home alone had collapsed to two tiers. So `--muted` goes **70% → 76% sumi**, solved to restore the ORIGINAL step rather than picked: oklab L ΔL to faint ratio on coast vs --ink was 70 / 62 .4884 .0558 1.269 4.95 2.299 now **76 / 68** .4492 **.0533** **1.259** **5.86** 1.941 ΔL lands within **4.5%** of what it was, and every other number goes UP: `--muted` on the worst ground 4.95 → 5.86. The one number that falls is `--muted` against `--ink` (2.30 → 1.94), which is not a contrast requirement — they are never adjacent as small type on each other, and ΔL between them is 0.167, an enormous step. 🚫 **Do not move one of these two without the other.** They are a ramp, not two values. ⚠ Declared on `[data-skin]`, which is `` — so `.band--ink`'s own re-point (§23.1) and `.band--sink`/`.reply-ask`'s 66% (the audit fix) are untouched: those set the property on THEMSELVES and their subtrees inherit it. Different elements, no cascade conflict. ⚠ A band with an OPAQUE background hides the backdrop completely, which is why those two registers did not need re-measuring against it. */ [data-skin="hanakenji"] { --muted: color-mix(in oklab, var(--sumi) 76%, var(--paper)); --faint: color-mix(in oklab, var(--sumi) 68%, var(--paper)); } /* ═══ §40 · ROUND TWENTY, PART THREE — THE WHOLE GATE (2026-09-17) ════════════════════ The user's note: one gate image per picture aspect, **the photograph sitting fully INSIDE the gate**. That retires the whole crown+post pipeline on these two plates — §19.4's 3-slice, §26's four constants and §37.3's 7-slice re-cut — by no longer matching. Nothing is deleted: `torii-crown-red.png` / `torii-post-red.png` stay because §24.4's rail mask still reads `torii-gate-red.png`, and §37.3 remains the record of how they were cut. ── what arrived, and what the lab had to do ───────────────────────────────── The brief's contract was on the OPENING — the clear area between the posts had to be the picture's aspect at top .24 / bottom .06 / sides .11, so file = opening × (1.22 w, 1.30 h). **Round one of the gates missed it on two of three** (`IMAGE-QA-round20.md` has the numbers: openings of 0.98 and 0.45 against 1.333 and 0.800, and gates **1.87×** and **2.64×** their own opening — an 878px gate around a 390px plate). The corrected pair fixed the disqualifying half: file opening aspect needed gate ÷ opening torii-gate-43 1.093 1.333 **1.535** torii-gate-169 1.756 1.778 **1.532** (accepted first time) torii-gate-45 0.588 0.800 **1.589** ⚠ **The opening aspect is still not the picture's, and it is ACCEPTED rather than unnoticed.** The consequence is only that the posts stand a little wider of the picture than briefed — the gate is placed off its opening's WIDTH and its nuki's UNDERSIDE, so a taller opening simply means more post below, and the lab cut that to a uniform hairline (below). ⚠ **`gate ÷ opening` is the number that decides a gate**, and ~1.53 is the shipped value on all three — which is not a coincidence, it is §19.4's own geometry: the retired crown drew at **150%** of the plate and the band has clipped its tips ever since. So this route inherits a clearance problem the page already solved. 🚫 A future gate over ~1.6 does not fit. ⚠ **All three were cropped to content with their legs cut to 1.0% of the picture's height.** As delivered the posts ran to the file's bottom edge, which is 22–36% of the picture's height of free-standing leg — and unequal between the three, which on Travel would have been two different gates on one page at two widths. One uniform cut is what makes them a family. ── the geometry, measured off the cropped files ────────────────────────────── `.torii` is `inset: 0` on `.torii-frame` (§19.4), so with the frame's padding dropped its box IS the picture. The gate is then three numbers, all fractions of the PICTURE: --sk-gate-w the gate's width, as % of the picture's width --sk-gate-x the opening's left edge, negated — % of the picture's WIDTH --sk-gate-y the nuki's underside, negated — % of the picture's HEIGHT (`top` resolves against height, so this one is NOT the same denominator as the other two) --sk-gate-lift the figure's own `margin-top`, to clear the lintel. ⚠ A percentage margin resolves against **WIDTH**, so it is `--sk-gate-y × (picture h / picture w)`: 43 → 40.24 × .75 = **30.18%** · 169 → 51.12 × .5625 = **28.76%** · 45 → 27.55 × 1.25 = **34.44%**. Re-derive it, never copy it between plates. ⚠ **`--sk-gate-y` differs wildly between the three (27–51%) and that is correct**: it is the kasagi-to-nuki gap measured against three different picture heights, not a style choice. ── the traps this route walks into ────────────────────────────────────── ⚠ **`max-width: none` is load-bearing** — `engine.css` caps every img at 100% of its container, which would draw the gate at the picture's width and put the posts inside the photograph. §19.4 paid for this once already. ⚠ **The selector reaches through `