nightly: re-record differential.json so #1384's twelve new check kinds can alert #1395

#1384 (merged as 29bfc85) landed seven differential pull requests. Five of them add check kinds that parity_checks/nightly/baselines/differential.json does not have yet. A row missing from the baseline is reported as new, and a new row never alerts. Until the baseline records these rows, a regression in any of the new checks goes unreported.

On #1384's own nightly (run 40904665252) the new kinds add 14,066 rows: 28,949 in all, against the baseline's 14,883, with none missing.

#1317's generated models are not part of the nightly. #1323's NF arm has its own baseline, differential_nf.json, committed in that PR.

Gated on: a scheduled nightly on main at or after 29bfc85. The next one is 2026-10-30 08:17 UTC. A workflow_dispatch of nightly-parity on main gives one sooner.

To do:

1. Take that run's nightly-differential artifact (verdicts.json) and diff it against the committed baseline with parity_checks/nightly/verdicts.py. Expect the twelve kinds as new, and no alert.
2. Before recording them, look at the new kinds' DIFF and EXCEPTION rows. Recording one makes it a known failure, so each needs a reason in the pull request, or an issue of its own. On run 40904665252 there were 17:
   • reorder: BIOMD0000000908 DIFF, 1.09e+134 of the trajectory's peak with species, reactions and parameters reversed. MODEL1812040005 EXCEPTION (CV_REPTD_RHSFUNC_ERR at t = 1).
   • restart: BIOMD0000000827 EXCEPTION (CV_CONV_FAILURE at t = 106).
   • stable: 10 DIFF, each "contested" by differential: attribute a disagreement with a third arm, and flag answers that flip with the grid or tolerance #1309's design (BIOMD0000000620, 621, 755, 1043, 1060; MODEL1109160000, MODEL1911100005, MODEL1912090001, MODEL2310250001, MODEL7908934508). 4 EXCEPTION (BIOMD0000000336, 763, 810; MODEL1006230034).
3. Copy verdicts.json over the baseline in a re-baseline pull request. Baseline-only pull requests skip the nightly sweep by the paths filter.

Overlap with #1371. #1371 re-records the same file for the four klu_vs_dense rows #1295 fixed, and is gated on the same kind of run. One re-baseline pull request from one nightly can close both.

Done when the re-baseline pull request is merged and the next nightly reads the twelve kinds as unchanged rather than new.