The generated headline said mdsmith is "closing the rest of that
gap" to the Rust linters, implying it trails them. Full mdsmith
(214/154 ms) now beats rumdl and panache and trails only mado,
so the gap is to the fastest of them, not to the group.

Fix the hardcoded narrative in gen_fragments.py and regenerate
headline.fragment.md plus the performance.md include. The results
table is byte-identical (same data), so bench-fragments stays
green. `mdsmith check .` passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThu9DcNDDSAKaw1upanY4

Statement-by-statement validation against the data, conventions,
and code surfaced three errors:

- "gomarklint's 20-rule default set" -> 21. gomarklint ships 23
  rules with 2 off by default (external-link, max-line-length),
  and gomarklint-parity resolves to 43-25+3 = 21 mdsmith rules.
- Residual asymmetries claimed MDS031 unclosed-code-block is
  "not in the parity disable list" and "kept on". It is disabled
  in mado-parity, rumdl-parity, and markdownlint-parity (the
  embedded fragment lists it), and mdsmith-parity now selects
  mado-parity. Reworded: the slip-through risk is the two opt-in
  rules MDS034 and MDS067; MDS031 is no longer an exception.
- check-bench budgets/baselines were stale. bench_test.go sets
  Small = 250 ms budget / ~14 ms baseline and Large = 2 s budget
  / ~90 ms baseline, not "2 s / 12 s" and "~0.09 s / ~0.8 s".

`mdsmith check .` passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThu9DcNDDSAKaw1upanY4

Adversarial cross-statement validation found the doc told two
different "fastest tool" stories and a backwards parity claim:

- "gomarklint is the fastest tool in the benchmark" (unscoped)
  vs "Pick mado for the fastest..." vs the committed table where
  mado is fastest and gomarklint has no row. Scope every
  gomarklint "fastest" to the per-merge copy / lightest workload;
  frame mado as fastest at a full markdownlint rule set and
  gomarklint as faster-but-checks-least. Headline and fairness
  note already agree.
- "gomarklint-parity runs no anchor check at all" is backwards:
  convention.go keeps same-file-anchor (MDS070) enabled in
  gomarklint-parity (plan 2606210840); only mado-parity disables
  it. Rewritten to say the parse-skip-safe anchor check stays in.
- "more than an order of magnitude behind every native tool" is
  false for panache (cli2 is ~6x, not >10x, behind it); "all
  finish in tens of milliseconds" is false (mdsmith/rumdl/panache
  are hundreds). Scoped both.
- Dropped the "21-rule" count that clashed with the "22" coverage
  figure; the set is now described as "enables 3, disables 25".
- markdown-linters.md: scope "mado leads the per-file race" to
  the Rust linters; note parity trails mado on the neutral corpus.

`mdsmith check .` passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThu9DcNDDSAKaw1upanY4

On the parity workload mado and mdsmith-parity tie (67 ms = 67 ms
on repo; 47 vs 53, within noise, on neutral), so recommending one
over the other "for throughput" implies a difference that is not
there. Reframe "Which tool to pick": the fast tools cluster
within run-to-run noise (mado, mdsmith-parity, with gomarklint
quicker because it runs the fewest rules); only rumdl and panache
clearly trail. Choose on capability, not on these margins.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThu9DcNDDSAKaw1upanY4

The committed data predates the per-linter parity split, so it
carries a single `mdsmith-parity` row — but the generated results
note described per-linter `mdsmith-<linter>-parity` rows that
aren't in the table, which read as a contradiction.

Make gen_fragments.py emit the note that matches the rows the
data actually holds: the single pre-split column now, and the
per-peer wording automatically once a refresh adds the four
columns. Align the "Like-for-like" parenthetical to the same
framing (single column reads closest to mado-parity; the next
refresh replaces it with one column per peer).

Regenerated; results table byte-identical; `mdsmith check .`
passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThu9DcNDDSAKaw1upanY4