tools/zig: one manifest → kernel set step for every family, reading a Recorder dump as well as a manifest #792

What exists

• tools/zig/triton_aot_manifest.py records a run's launches (Recorder.dump → launches.json) and maps each
  launched specialisation to its cached cubin (build → manifest.json).
• zig/tests/cuda/nemotron/aot_pack.py turns one or more manifests into the set the engine loads: aot.json in the
  KernelJson schema (zig/src/cuda/aot.zig:32-46) plus cubins/<hash>.cubin. It is documented only in step 3 of
  zig/tests/cuda/nemotron/box/README.md. Qwen3.8-27B on CUDA: the qwen27 family's engine for NVIDIA GPUs, every stream drafted #608 packs the 27B's set with it as well.
• packaging/README.md (-Ddist-cuda-aot) and RUNBOOK.md (TENSORFOLD_CUDA_KERNELS) describe the set's layout,
  sm<capability>/aot.json and cubins/<captured-hash>.cubin, but not how a set is made.

What is missing

1. Every CUDA family's set goes through this step, but it is filed under one family's tests, and the docs that
   describe a set do not point at it.
2. It reads only the manifest shape: kernels as a list, cubin relative to the cache, abi, cubin_sha256. A
   Recorder dump has a different shape: kernels is a mapping keyed by hash, params lists every parameter
   including the constexprs, and the cubin is an absolute path in files['<name>.cubin']. The 0.6.6 capture made on a
   GB10 for the 27B came in that shape (from a newer copy of the Recorder, keyed by cubin sha256), and it cannot be
   packed without first running build against the same cache.
3. --jit is required, and only the in-tree capture scripts write jit.json. A capture made outside the tree (the
   10 Oct rule, see Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090) #726) has none.
4. --cache must be repeated beside every --manifest, although each manifest records its own cache
   (triton_aot_manifest.py:265).

Proposal

Move aot_pack.py to tools/zig/triton_aot_pack.py, beside the manifest step, and change it as follows:

• Accept either shape for each --manifest. The runtime arguments are the signature entries whose type is not
  constexpr, in params order when the input has params.
• Find each cubin by cubin under the cache, or by files['<name>.cubin']. When that absolute path was the run's own
  (a container mount), look for it under --cache.
• Make --jit optional. Without it no parameter is marked nospec. By matches(), that can only refuse a launch
  (an int argument of 1, or a multiple of 16, that Triton was told not to specialise: aot.zig:191-192). It can
  never serve the launch a variant compiled for other values.
• Default --cache to the manifest's own.
• Write a constexpr only when its name is not a runtime parameter. The matcher reads a variant's constant for an
  argument only when that argument is not one of the variant's params (aot.zig:183-187).

These stay as they are: the fields written, the PTX check wherever the input has abi, the sha256 check wherever it
has cubin_sha256, and the sort order. The Nemotron box README and test_aot_manifest.py follow the file to its new
path.

Two rules any manifest → set step must keep

1. Triton's hidden launch pointers are not params. Triton's launcher appends global_scratch and
   profile_scratch; the caller never names them. matches() ends with return runtime == k.params.len
   (aot.zig:209), so a variant that lists either one can never match. aot_pack.py already leaves them out, and it
   refuses a cubin whose PTX parameters are not the runtime arguments followed by exactly those two
   (aot_pack.py:54-57). The new file keeps that check wherever the input has abi, and it builds params from the
   signature for both shapes. A set whose kernels need scratch is refused at load (aot.zig:100).
2. div16 records what Triton proved, not what the caller will pass. The step writes it as captured: a
   tt.divisibility of 16 in attrs. Whether a variant without it may serve an aligned pointer is up to the
   matcher. Main refuses it (aot.zig:196, strict equality); Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090) #726 (ff7cc22) accepts it in a second pass. See
   aot.zig: a captured variant compiled without 16-divisibility refuses an aligned launch #791.

Effect on existing sets

None. The engine reads kernels and ignores every other field (aot.zig:91). Sets already packed are not touched,
and tools/release/package.sh copies sets without calling this step. For input that aot_pack.py accepts today, the
output differs only in three places:

• generator.
• div16, where an argument carries an attribute other than tt.divisibility 16. aot_pack.py:53 counts any
  attribute as divisibility.
• Constants that share a runtime parameter's name, which are no longer written.

Gate: python3 -B zig/tests/cuda/nemotron/test_aot_manifest.py, run against the new path and given a second case
that packs a Recorder dump directly. It is host-only and needs no GPU or model.