Round 3 · still swap

Still-swap agent · 2026-10-13 · files changed: prototype/mind.html (the #head slot and its CSS) and prototype/standard.html (the .hero-head slot and its CSS). Nothing else was edited: index.html, shared.css, server.py, img/, data/, brief/ and research/ are untouched, and no earlier shot or report was overwritten.

Method: headless Chrome for Testing (chromium-1234) through the Playwright library, 1440x900 at DPR 2 and 390x844 at DPR 3 with isMobile and hasTouch, served by python3 -I -m http.server 8705 --bind 127.0.0.1 from prototype/ (PID 77660, stopped by that PID at the end, port 8705 confirmed free; 8777 was never listening and never contacted). Both pages were loaded with ?demo=1, so neither page calls /api/; a context route aborted and counted every request to port 8777 or to /api/, and both counters ended at 0. Shot prefix r3-still-. Every capture cited below was read as pixels before it was judged. Analysis ran through uv run --script in scratch (r3-still), removed at the end.

Summary

1. The stills (step 1)

Measured on the files (alpha > 0 extent, uv run --script):

Darkest opaque pixel L 170.4 / 170.2 (the opening's shading floor), widest row at 55 % of the height. Read as pixels: both files show the round-3 head (shallow soft dishes well inside the silhouette, brow, bridge, white dissolve) and match the head in r3-opening-1440x900.png and r3-opening-head-1440.png. Not a skull, helmet, capsule, ball with holes, goggles or sleep mask.

2. The swap (step 2)

Markup

In both slots the inline 240x200 SVG (with its private mSkin, mHollow, mFade, mMask, mClip, mBlur, mBlur2 defs, which nothing else referenced; grep across all four prototype files) is replaced by one element inside the unchanged wrapper:

<img src="img/head-still.png" srcset="img/head-still@2x.png 2x" width="800" height="924" alt="" decoding="async" draggable="false">

CSS: #head img and .hero-head img are display:block; width:100%; height:auto; margin-bottom:-3.3%. The width/height attributes only supply the aspect ratio before decode, so the slot (and the standard mask) has its final height from the first frame. The wrappers keep aria-hidden (standard) or sit in the aria-hidden realm (mind); alt="" makes the image decorative either way. The gHead symbol and its three <use> sites (mind top bar, dock trigger, dock head) are untouched.

The -3.3% foot: percent margins resolve against the slot's width, and the still's transparent foot is 3.375 % (1x) / 3.31 % (2x) of its width, so the slot's box ends where the dissolve ends. Measured on the frozen captures: the last row the still changes by 1 level or more is 237.0 against a slot bottom of 237.0 (mind 1440) and 237.7 against 238.3 (mind 390); on the standard hero the slot bottom is the hero's lower edge (868 at 1440) and the mask plus dissolve fall under 1 level by 857.5.

Geometry: why the slot widths changed

The SVG slot was 0.8333 x its width tall. The still, once its foot is trimmed, is 1.122 x its width. At the old widths every slot would have grown 35 % taller: on the mind page the head would have ended at about 300 px, on top of the opener's h1 (272) at 1440 and at 390, and on the phone hero 71 % of the head would have gone behind the copy. The smallest change that keeps everything the mind-standard fixer measured valid is to keep each slot's height, top, bottom and anchor and narrow its width by 0.8333 / 1.122 = 0.743:

Both clamp ends scale by the same factor, so at every viewport each new box has the same top, bottom and height as the old one (to within 0.3 px at the rounded clamp ends) and is narrower around the same anchor (centre for mind and phone, right 5 % for the desktop hero): it sits inside the old box. Nothing that cleared the SVG can collide with the still. The phone hero's padding-top: calc(var(--head-w) * .3333) and --head-w: min(62vw,260px) are unchanged and still mean what their comments say (the token's comment now reads "the phone head box: .8333 x --head-w tall (the still at .743 x --head-w wide)"); the slot comments in both pages explain the factor and the foot.

Resulting heads (footprint = pixels the still changes by 1 level or more, frozen captures):

Kept as built

The standard masks (55 % desktop, 64 % phone), the phone copy's white gradient, the mind slot's ::after contact shadow, both breathe animations and their transform origins, the reduced-motion rules, the mind fade (body:not([data-state="empty"]) #head), the dock-open fade and its max-width 820 reset, and the .facts rules. Diff against the pre-swap files: width values, the img rules replacing the dead svg rules, the slot markup, and comments; nothing else.

3. Verification (step 3)

Captures (all in prototype/shots/, read as pixels)

One thing visible only in the stretched crown panel: a faint horizontal contour at CSS y 64-68 above the crown. It is identical with and without the still (row means differ by 0.00): the top bar's fade and the realm glow's banding, not the head.

Edge checks (frozen pairs)

Method: a second load of each page, every animation paused at t = 0 (document.getAnimations()), captured with the still and again with img{visibility:hidden}; dL is the luminance difference. Bands are 3 device px just inside each slot edge and 4 px just outside it; the ring is the 4 device px just inside the head's footprint in the top 80 % of the slot (above the dissolve), over a light background (text excluded).

• No rectangle: the still changes nothing outside its slot box and nothing along the left, right or top edge of it. The 1-2 levels in the mind bottom band are the last rows of the dissolve itself, which ends exactly there (the band below it is 0).
• No halo: the lighter ring pixels sit on the crown, where the key light makes the still's own edge bright (top-edge partial pixels median straight L 244-245 against 203-207 on the sides). In the still, every partially covered edge pixel matches the opaque pixels within 3 px (1x: 5th to 95th percentile -3.9 to +4.3 levels, max +8.3; 2x: -2.5 to +3.4, max +9.0), so there is no white or dark matte fringe to composite. Composited, the crown is at most 4 levels lighter than the fog around it.
• No dark sockets: the darkest head pixel is 173.4 on the mind page (the still's floor 170.2 at opacity .96) and lighter under the standard masks.

Console and network

Eight loads (mind and standard, normal and frozen, at both sizes): 0 console errors, 0 page errors, 0 failed requests, 0 responses of 400 or more, 0 requests to port 8777, 0 requests to /api/. currentSrc is img/head-still@2x.png at DPR 2 and 3 (naturalWidth reports the density-corrected 800). Server log: 61 requests, all 200; the 1x file was fetched only by the curl check, the browser always took the 2x.

4. DOM before and after (slot box incl. the breathe transform, CSS px)

Opener h1 (mind) 272 at both sizes and the hero copy, eyebrow, mark, theme and facts boxes (standard) are identical before and after.

5. Not done, notes

• Only the requested states were captured. Other viewports were not screenshotted; the containment argument in section 2 (same height and anchors at both clamp ends, narrower box) covers them, and the dock-open hero at desktop widths fades the head anyway.
• Reduced motion, no-JS and the phone dock were not re-captured; their rules were not touched (the still also renders without JS, as the SVG did).
• decoding="async" can show the slot one frame before the head on a cold load; the width/height attributes keep the slot's size and mask right meanwhile.
• The mind head is now 138 px wide at 1440 (the SVG capsule was 165). That is the design's 38 % of the opening scale and the slot's old height; if the owner wants it larger, the opener has to move down first (stage padding), which is outside this agent's slot-only remit.
• Cleanup: static server PID 77660 stopped by that PID, lsof shows 8705 free; scratch r3-still removed; backups of the pre-swap pages lived only in that scratch directory.