Batch chunk requests and add region level-of-detail

[gh8sted]

Chunk requests and responses were one packet each. A client loading a screen at low zoom sent thousands of 8-byte requests and got thousands of separate messages back. This batches both directions and adds a coarse region format so a zoomed-out view can be painted before full detail arrives.

Requires the matching client change: OurSources/owop-client#NN. The protocol changes on both sides — neither half works alone.

Batched chunk requests

New request packet: [u16 guard][u16 count][u16 reserved] followed by count × (i32 x, i32 y). The length is always 6 + 8*count, which no existing packet length can equal, so it is checked ahead of the length-based switch without disturbing any existing case.

Response 0x0B carries [u16 count] followed by count × ([u16 len][0x02 packet]), where each embedded packet is byte-identical to a standalone one.

Chunks in regions that aren't loaded yet turned out to be the important case. Those go through awaitRegionLoad, which sent one message per chunk, so batching the request alone changed almost nothing in practice — on a cold client 81% of chunks (3,328 of 4,096) still arrived individually. Consecutive deferred chunk loads are now gathered into one message, flushed before any other action so ordering relative to pastes and erases is unchanged.

Measured on a cold client over a 64×64 chunk grid: 157ms → 90ms, with every chunk batched instead of 3,328 individual messages. An 8,100-chunk view previously never completed at all, stalling at roughly 5,800; it now finishes in 90ms across 50 messages.

Region level-of-detail (0x0C)

One averaged colour per chunk, so a whole region is 768 bytes instead of the ~200KB its 256 chunks cost at full detail. The request is a fixed 16 bytes describing a rectangle of regions. The downsample is cached on the Region and invalidated in flagDataModified. Cold regions defer exactly like chunk requests do.

Verified against full detail: 16,384 chunk averages compared, 0 mismatches, maximum channel error 0.

Limits

• Requests are capped at 4,095 chunks and 1,024 regions — past maxPayloadLength a frame cannot arrive intact anyway
• Response messages are capped at 512KB (chunks) and 256KB (LOD) so a single request cannot force a multi-megabyte send

Notes

tools/fillWorld.js is a development utility for generating large test worlds. It is in its own commit and can be dropped if you would rather not carry it.

Two caveats on the numbers above, both worth knowing before reading too much into them:

• Transfer figures come from a test world of per-pixel random noise, which defeats the run-length encoding completely (~780 bytes per chunk). Real world data compresses far better, so the byte counts here are pessimistic.
• LOD's benefit scales with view size. On a small cold area it is only about 1.4× faster, because both paths still load the same regions from disk and that dominates. It eliminates network transfer and client-side decode, not disk I/O.