[kotevcode]

What this changes

The CUDA kept-state identity (cuda_disk.identityOf) is now written once, by components(h, …), into any hasher with update. identityOf passes Wyhash and keeps every kept key byte for byte; the execution witness (#2741) will pass SHA-256 for the same list, so the two can never drift. Behaviour is unchanged.

• components takes a Components struct; the binary, checkpoint stamps and probe are optional, since the witness binds the binary and the weight bytes itself. Kept keys always pass all three, in the old order.

• disk_identity and file_identity helpers take any hasher; file_identity.running() returns the running binary's digest and bind writes it as current did. aot_identity.write feeds a captured cubin set to any hasher; hash wraps it in Wyhash as before.

• New cuda_disk_golden_test.zig: main's identity sequence, copied update for update, must equal identityOf over this binary, a real checkpoint and a captured set; and a fixed-input key computed on unmodified main is pinned (skipped when the build embeds real kernel images). Swapping two components fails both. A SHA-256 case checks a changed cubin, profile value or device changes the result.

• Issue: Refs Opt-in coherent execution witness for exact-response cache qualification #2741. The behaviour-free step agreed there to go ahead of the feature.

• Stacked on: main

• atc overlap --json: run on this head against main; complete: false (otherPullRequestsLimit, pullFilesUnavailable, timeLimit). 7 pull requests report a textual conflict, each only in CHANGELOG.md: Vulkan: qmm_q4_vec4 reads the 4-bit projection 16 bytes at a time, 34.8% to 43.0% less run time on an M1 with the same logits (stacked on #1544) #1990, dsv41 metal: image prompts #1722, dsv41 metal: the vision tower's kernels under the kernel contract #1721, dsv41 metal: the lane target, the prompt cache and the native server over Metal #1720, Implement GLP control vectors with a fused Metal kernel for Qwen3.8-27B #1835, app: control vector picker and strength in Engine advanced #1951 and GLM-5.3-Flash on CUDA keeps conversation states on disk, every rank's part (--learn-conversations) #1411. Native Vulkan backend for Qwen3.5-2B (opt-in): engine and kernel-contract registration (stacked on #1526) #1544 and Unlimited-OCR-2 on native Metal: the publisher's BF16 release, page images to structured text #2074 are clean; 4 report unknown.

Checks on this head

• Head commit: 7795508
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 345 and 345, 0 new, after rebasing onto main eab305d (only CHANGELOG.md conflicted). python3 tools/pr_check.py --base upstream/main --pr 2903: 0 errors.
• Tests run, with the command and pass, fail and skip counts (macOS 26.6, M5 Max):
  • zig build test-cuda-host: 379 of 379 steps, 6149 passed, 0 failed, 38 skipped.
  • zig build test-disk-identity-host: 34 passed. zig build test-cuda-persistence-source-host: 4 passed. zig build test-cuda-disk-open-host: passed.
  • zig build compile-disk-identity-metal-host and compile-persistence-host compile.
  • The golden tests fail when two components swap order (checked by planting the swap).
  • zig fmt --check build.zig zig and git diff --check pass.
• CI: not yet run on this head.
• Not run, and why: a kept state written on a GPU by main and reopened by this branch; the pinned key and the update-for-update copy cover the bytes on the host. Metal steps; this Mac has no Xcode Metal toolchain.

Receipt

n/a: the identity's input bytes and their order are unchanged, so every kept key is the same and nothing that reaches inference changes.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (They pin unchanged behaviour; a planted reorder fails them.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

The inference checklist is skipped because the Receipt is n/a.

[ashhart]

Thanks @kotevcode, this is the behaviour-free step we agreed on in #2741, and it is clean.

I went through identityOf update by update against main. The binary digest, checkpoint, kernel images, captured set, GPU name, capability and driver, facts, probe and profile values go into Wyhash in the same order, with the same widths, and nothing optional is ever left out on the kept-key path. The golden test's reference is main's body, and I reproduced the pinned key (9113307822166948912) independently from main's byte sequence, so it really is main's key. Skipping that one case when kernel images are embedded is fine: those bytes are build-specific, and the update-for-update comparison still runs there.

No asks. It's queued to land after today's release once the remaining CI checks on this head are green.

A note for #2901, not this PR: captured() chooses the 64-bit fold only when the hasher is exactly std.hash.Wyhash, so a wrapper hasher would take the SHA-256 branch. Keep that in mind if the witness ever passes something other than SHA-256.

• Synthena (Ash's Agent)

[ashhart]

Thank you @kotevcode. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @kotevcode! This is in TensorFold 1.2.8, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.8

• Synthena (Ash's Agent)