model-matrix.md prose counters drifted: LTX-2.5 was added to the rows and the rollup but to none of the four sentences that count them

[localai-bot]

cefacd2d0 (#641) added LTX2VideoTransformer3DModel to .agents/model-matrix.md
and bumped the CI-enforced surfaces — the MODEL pin (369 -> 370) and the rollup
table — but left every prose counter in the file untouched. Those counters
are enforced by nothing, so they drift silently.

Measured against the file itself (scripts/check-agent-record.py's own parser
over .agents/model-matrix.md: 370 MODEL rows, 321 INVENTORIED, 49 engaged):

Claim in the file	Said	Actual
"plus 10 rows that the pinned registry does not contain"	10	11 — LTX-2.5 is BEYOND-PIN AND OUT-OF-REPO and is not in the enumeration
"326 architecture rows are inventoried at the pin ... = 336 total"	326 / 336	355 / 366
"Of the 336, 45 are past INVENTORIED, the remaining 291 ..."	45 / 291	49 / 321 (of 370 rows)
"Engaged architectures (the 48 non-INVENTORIED rows)"	48	49
"NOT part of the 328 architecture rows counted above" (## MODEL-AUDIO)	328	366

Two separate drifts are tangled here. LTX-2.5 is the immediate one: the
not-at-the-pin enumeration and the engaged count both moved by exactly one row.
The 326 / 336 / 45 / 291 family is older and was already inconsistent with
the rollup before #641 — eba6ab7c7 (#625) refreshed the enumeration from 3 to
10 and the rollup from 314 to 321, but carried 45 and re-derived 291 from the
stale 336, so the paragraph has been describing a population no reader can
reproduce for several landings.

The not-at-the-pin population is checkable: exactly 11 rows declare in their own
Upstream cell that the pinned registry does not contain them (BEYOND-PIN,
OUT-OF-REPO, TARGET-PENDING, pin-lag), and the ten already named plus
MODEL-DIFFUSION-ltx-2-5-... are those 11. The residual 4 rows between the 366
architecture rows and the rollup's 370 are the ones that say in their own
sections that they are not architectures: MODEL-FACTORY-registry, the two
MODEL-AUDIO component rows, and the single MODEL-HFDYNAMIC row.

scripts/check-agent-record.py pins the MODEL row COUNT and
scripts/check-model-checklist.py pins the rollup against the real per-state
row counts, so both stayed green throughout. Neither reads prose, which is
exactly the gap.

Repair here is the recount plus a pointer to how to re-derive it, so the next
editor recounts instead of carrying a number forward. Whether these hand-kept
prose counters should exist at all — they are the same "a number that moves on
every edit" shape AGENTS.md warns about, and #622 records the NOW.md instance —
is a separate question this does not decide.

Records only.