check-public-doc-tables: an EMPHASIS-lead paragraph is counted by nothing, so a **bold**-led prose wall is unbounded on BENCHMARKS/FEATURES/STATUS

[localai-bot]

Found while repairing review finding F1 on #494 (row ENG-RECORD-CONFLICT-SURFACES, spec .agents/specs/benchmarks-entry-cap.md). Filed, not fixed there, because fixing it owes an edit to a page another PR holds open.

The defect

_prose_paragraphs in scripts/check-public-doc-tables.py excludes any line beginning with *. That was meant to exclude list items. It also excludes every paragraph that OPENS with emphasis, which is the house style on these pages:

**Protocol.** Greedy, closed loop, three interleaved repetitions per point ...

Such a paragraph is counted by nothing: not max_prose_paragraphs, not MAX_PARAGRAPH_CHARS, not MAX_CELL_CHARS, not MAX_ROW_CHARS, and not DATED_HEADING_RE. #494 closed the sibling channel (bullets and blockquotes are now folded into the paragraph running through them) and left this one open, with the residue recorded in the function's docstring and pinned by test_an_EMPHASIS_lead_wall_is_a_KNOWN_residue.

Reproduce on current main or on #494's head:

wall = "\n\n".join(f"**Attempt {i}.** " + "x " * 200 for i in range(500))
doc_tables.benchmarks_errors(VALID + "\n\n" + wall)   # -> []

Why #494 did not fix it

Detecting a list item precisely (- , * , + , 1. ) instead of by bare prefix makes emphasis-lead lines prose, which is correct, and immediately turns FOUR paragraphs already shipped on docs/BENCHMARKS.md red against MAX_PARAGRAPH_CHARS = 700, at 717, 719, 748 and 1,084 characters (around the 35B-by-concurrency and How-we-measure sections). Closing the hole therefore requires editing the page, and #481 rewrites the 35B row in place. #494 is scoped to leave docs/BENCHMARKS.md unchanged for exactly that reason.

It also changes what an existing gate counts, which per AGENTS.md takes its own spec and its own red-before rather than riding along.

What closing it takes

1. Replace the bare * prefix exclusion with LIST_ITEM_RE (already defined in the checker) so emphasis-lead lines are prose.
2. Split or shorten the four over-budget paragraphs on docs/BENCHMARKS.md, coordinated with gate(#378): the first DEVICE run of the warp MoE router, and two record corrections #481.
3. Re-baseline max_prose_paragraphs on both pages and STATUS_RATCHET["long_paragraphs"] to what they then measure.
4. RED-before: test_an_EMPHASIS_lead_wall_is_a_KNOWN_residue in tests/scripts/test_check_public_doc_tables.py goes red when this lands, and should be inverted into a passing rejection test in the same change.

Recorded as W8 in .agents/specs/benchmarks-entry-cap.md.