[apresmoi]

Change

On phones the Milky Way flickered (layers blinking out every other frame) and on the iPad it froze for about 200 ms when crossing into or out of the galaxy. Both came from the far galaxy drawing 390 full-screen layers: 52/52/26 slabs, each drawn three times as optical copies.

• Phone (Galaxy S23, Chrome): Chrome caps tile memory at 256 MB. While the galaxy turned, it ran short in 827 of 867 tile assignments and drew layers empty in 248 frames of an 8 s turn. That was the flicker.
• iPad: adding or removing the bank cost 191, 185 and 214 ms composites.

What changes:

• The far bank is 13/13/7 slabs, and each slab image already holds its three optical copies (alpha 1 − (1 − a)³). It's drawn as one layer per slab at opacity gain / 3, so there are at most 33 layers instead of 390. The bake gets an opt-in --merged-copies flag. The prepared volume declares mergedOpticalCopies: 3 and the bank's prototype records it, so a reused bank keeps it. Volume topology now tells merged and unmerged encodings apart.
• The far picture leaves as soon as the inside sky covers it (reverting the "wait until the camera stops" deferral from fix(milky-way): stop dropping frames when zooming into the galaxy #1586). Keeping 390 layers under the sky was what made the phone flicker while zooming in.

Body guide: Milky Way volume README.

Known limit, accepted by the owner on the phone: a merged copy is exact only for faint material. Face-on, an opaque pixel draws at a third and a half-opaque one at 0.29 instead of 0.5. That is 31% of the slabs' total alpha face-on and 0.8% at grazing views. The alternatives were measured and rejected: two layers per slab ran the phone out of tile memory again (56 to 62 frames with missing tiles), and fewer, thicker slabs (7/7/4) showed gaps between the sheets.

Prepared assets

• Baked the object (node packages/bake/authoring/milky-way-dust/prepare-layers.mts --slabs-only). A second bake reproduced the inventory exactly, and CSSEARTH_GALAXY_FULL_REPLAY=1 rebuilds all 33 slabs and 6 sky faces from the restored inputs byte for byte.
• Published the bytes (node packages/bake/cli/publish-runtime-assets.mts --object=milky-way-volume); check-assets-published verifies every key live.
• Committed the refreshed inventory.json, and no baked files.

Evidence and limits

Proves the change, same 8 s touch turn of the galaxy on the phone (Chrome traces):

Far galaxy	Tile memory short	Frames with missing tiles	Dropped frames
52/52/26, three copies (main)	827 of 867	248	164
13/13/7, merged copies (this PR)	0 of 536, 0 of 547	9, 4	36, 62
Before #1581	0 of 506	0	43

• Bisect of the zoom flicker on the phone (frames whose brightness flips against both neighbours): before feat(milky-way): ship settled far layers and inside skies #1581 0; feat(milky-way): ship settled far layers and inside skies #1581 3 and 1; fix(milky-way): stop dropping frames when zooming into the galaxy #1586 before its far-picture deferral 1 and 0; with it 31 and 13; with it reverted 1 and 0.
• Slab-local fit against the same reference: relative L1 0.13/0.13/0.07, against 0.20/0.20/0.15 for the 52/52/26 bank.
• Mutation-checked tests:
  • a merged volume draws one layer per slab, exact for faint alpha, with the dense face-on values pinned;
  • the validator accepts exactly three;
  • topology separates the two encodings;
  • reuse keeps the declaration;
  • the alpha merge keeps color.

Clean: pnpm typecheck, eslint on the touched packages, galaxy bake tests with the CI-pinned NumPy 2.5.4 and Pillow 13.0.0, renderer volume and universe tests, objects volume tests (369 passed), check-prepared-delivery, check-no-variables, check-no-hashes, check-color-spelling, typescript-ownership.

The phone still drops about 5% of frames on this view, the same as before #1581. That remaining cost is the star labels: about 240 composited labels restyled every frame. It's separate work.