Closes NVIDIA#1090. `mlir-dialects.md` states its own freshness rule: At catalog SHA-256 `26a9cc13` (the stamp in every `ops/generated/` file header), the dialect holds 570 operations across 42 modules ... If the header stamp no longer starts with `26a9cc13`, the counts on this page predate the catalog you are reading. Nothing enforced that rule, so the rule that would expose stale counts was itself unchecked. The page failed its own test for weeks before NVIDIA#1052 (560 printed against 570 in the tree), and **it is failing again right now**: the catalog hashes to `00372dfe`, five hand-written ops and five catalog entries after that refresh landed. Two hand audits in two batches is the pattern this closes. ## The gate One file hashed, one page grepped, no cargo and no network. The hash is the one `cuda-intrinsics-gen` stamps into every generated file -- `sha256_bytes(catalog_json.as_bytes())` over the file as committed -- so a plain `sha256sum` reproduces it, and the page prints its first eight digits. Three checks rather than one, so each failure says which thing moved: - The generated sources must stamp the hash `catalog.json` actually has. Failing here means the generated outputs are stale, and the message says to run `cuda-intrinsics-gen -- generate` -- not to edit the book. - The page must print exactly one distinct stamp. It names it twice, once stating it and once in the rule that tells a reader what a mismatch means, so updating one and not the other leaves the rule contradicting the statement above it. - The stamps must agree, and the failure prints both sides plus the two commands that re-derive the counts the stamp guards. It refuses to report success from a check that found nothing: a page reworded so no stamp is left fails with that stated, rather than passing vacuously. That path needed the `|| true` on the capture rather than on a later pipeline -- under `pipefail` a grep matching nothing aborted the script before the diagnostic, turning a reworded page into a silent non-zero exit. ## The refresh that makes it pass Five numbers and the stamp, all re-derived from the tree: | claim | was | now | |---|--:|--:| | catalog stamp (x2) | `26a9cc13` | `00372dfe` | | total operations | 570 | 575 | | hand-written operations | 21 | 26 | | hand-written `wgmma` row | 9 | 14 | | catalog entries | 1010 | 1015 | Generated stays 549 across 35 modules and every area row is unchanged, because NVIDIA#1085's five new catalog entries are aliases onto existing structural ops -- exactly what the page's "several intrinsics can share one structural op" sentence predicts. The +5 hand-written are all `wgmma`, from NVIDIA#1075, NVIDIA#1077 and NVIDIA#1079. The `wgmma` row's description moves with its count, as in NVIDIA#1052: it said "the m64n64k16 bf16/f16 shapes", and the module now holds four shape/type pairs -- `m64n64k16` bf16 and f16, `m64n128k16` bf16 (NVIDIA#1079), `m64n64k8` tf32 (NVIDIA#1077). ## Verification - Four mutations, each failing with its own message: a stale stamp, only one of the two stamps updated, the page reworded so none is found, and an edited `catalog.json`. The healthy tree passes. - Every count re-checked against the tree: 26 + 549 = 575, both tables sum to their stated totals (7 rows to 26, 11 rows to 549), and each still lists every module on disk. - All eleven guard scripts pass, `check-spdx-headers.sh` accepts the new file, `cuda-intrinsics-gen -- check` reports outputs current, the workflow parses with all nine jobs on `ubuntu-latest`, and the book builds clean under `sphinx-build -W`. `just check-guards` should list this script too, one line beside the other ten. Its recipe is being restructured in NVIDIA#1102, so that line is left for a follow-up rather than conflicting; no guard enforces that list, so CI is unaffected either way. Signed-off-by: Vyncint Ng <115854244+vyncint@users.noreply.github.com>

Commits

Commits on Aug 21, 2026

ci(book): gate the dialect page's catalog stamp against catalog.json