Publication #2 cards: an unbounded build claim, no KV dtype, and stale upload comments

## Defect

Three provenance gaps on the publication #2 cards (live and byte-identical to `publication/nemotron-30b-a3b-fit16gib/README.md` and `publication/nemotron-30b-a3b-sensitivity-maps/README.md`).

1. Model card lines 64-67: "The serve test below ran build b10326, and that build or later loads this file." Two builds are recorded, b10326 and b10362. #410 records a b10573 load in a ticket comment. No record bounds the claim above that. The #484 rule for the target-runtimes page is that no row says "or later".
2. Model card lines 139-146 name no KV dtype for the serve test. The Gemma 4 card names f16. The 96.00 MiB over 16,384 cells and six layers is consistent with f16 and the card does not say so.
3. Model card lines 19-33 and maps card lines 16-17 carry HTML comments reading "the dataset repo does not exist yet". The dataset shipped 2026-08-22. The comments ship in the raw card.

## Fix

Bound the build claim to the recorded builds. State the KV dtype from the serve log. Delete the stale comments. #410 owns the Usage section on the same card. #409 changes the same card's F16 rows. Land after or with those.

Found on 2026-09-03 during the #484 checks (chart #481).

Start with publication/nemotron-30b-a3b-fit16gib/README.md and publication/nemotron-30b-a3b-sensitivity-maps/README.md, focusing on the cited lines and the recorded serve-log details. Bound the build statement to the recorded builds, state the serve-test KV dtype, and remove the stale dataset comments; verify the cards are byte-identical to the published versions and coordinate with #410 and #409.

Written by the indexing model from the issue text.