[ashhart]

Part of #1857: DeepSeek-V4.1 reads its pack's weight spans past the page cache, so a load no longer fills the cache it then has to release.

What changes

• Pack.read reads DeepSeek-V4.1's weight spans with O_DIRECT through core/direct_io, in at most 8 MiB of aligned staging, with exact strided gathers.
• core/direct_io now owns the TF_LOAD_DIRECT switch and the chunk size. TF_LOAD_DIRECT=0 reads through the page cache as before. A file system that refuses O_DIRECT, at open or on a read, falls back the same way.
• The prepared-image identity adds pack_read.zig, so existing prepared folders rebuild once, the same on every rank.

The pack hash and the prepared-image reads still go through the page cache, so this narrows #1857's minimum during load rather than fixing it outright.

Checks

• Host: the loaded bytes equal a real mapping's across large tensors, unaligned starts and tails, chunk seams, strided, overlapping and repeated runs, and 65 shards, in both TF_LOAD_DIRECT modes. The lean check finds no new problems.
• An independent review: LAND after one change, which this carries (the switch moved into core/direct_io, so it stands on main alone).
• Linux host run (aarch64, ext4, head c6d1c720): test-dsv41-direct-host in both TF_LOAD_DIRECT modes, test-dsv41-env-host and core/direct_io's tests all pass, including the two that only run on Linux. The first run on the previous head failed one: under O_DIRECT, a pack cut short inside a span re-read the same aligned tail and reported ReadFailed instead of end-of-file, so the pack reader never said ShortRead. c6d1c720 fixes that and adds a test for it. A Mac never takes this path (no O_DIRECT).
• Device: a load on a Linux box with a DeepSeek-V4.1 EXL3 pack, comparing load time, MemFree and Cached against main. None of our Linux boxes has the pack at the moment, so this waits for that run and doesn't land before it.

[ashhart]

Rebased onto today's main (c2090ff8c8). The changelog was the only conflict, and one commit dropped out because main already carries it. test-dsv41-direct-host passes in both TF_LOAD_DIRECT modes, with the environment and load-cache steps and the Linux DeepSeek server build. What's left is the GB10 run with the DeepSeek-V4.1-Flash EXL3 pack.

• Synthena (Ash's Agent)

[ashhart]

Device check status: blocked for now. On main, the CUDA server refuses DeepSeek-V4.1 behind its whole-model gate (#827). Its prompt path still needs captured Triton kernels, so a served main-vs-branch run can't start. The EXL3 2.9 bpw checkpoint is 197 GB, so it needs two GB10s at TP=2; one box can't hold it. This PR waits for #827, or for a two-GB10 run once both boxes have room for the checkpoint. Host tests and CI stand as they are.

• Synthena (Ash's Agent)