PR-2: pilot renders the full page, styled per D1 (−85.6% CSS) - #557

Part 1 (content): layouts/next/single.html ports the source page's frontmatter
sections - overview narrative, outcome stats, FAQs as a dl - 580 words vs the
live page's 587 (was 47). The tenure stat replicates the legacy template's
DERIVATION (now.Year - foundingYear), not the frozen frontmatter 18.

Part 2 (style): Artifact Editorial per the 4-1 competition vote. Paper
#fbf9f8, serif headline stack with system fallback (nothing fetched - repo
self-hosts only Caveat/Space Grotesk; self-hosting the D1 faces is its own
PR), ruby confined to hairlines, small-caps labels, the hero arrow-link and
ONE filled CTA. 0px corners, no shadows, 1px outlines.

Opus verifier returned a HOLD with 12 findings; all applied:

- F1 BLOCKER: `.next-article a` (0,1,1) outranked both single-class CTAs
  (0,1,0) - the page's one filled button shipped BLACK on ruby at 3.33:1, an
  AA failure. One-token fix: `:where(a)`. Re-verified in computed styles:
  button now white-on-ruby 5.13:1, link ruby.
- F2 BLOCKER: PR-1's near-empty baselines would fail every run - re-recorded
  via bin/record-baselines (kept 2, restored 54); Linux keys via CI dispatch
  on this branch.
- F3: the composition swap (FAQs in, clients/testimonials out - shared legacy
  partials, out of scope) dropped EVERY trust signal. Restored the minimal
  canon-exact proof line: "Rated 4.8/5 on Clutch" with the profile linked, no
  review count.
- F5: FAQ 5 carried unsourced "15+ years of industry experience" contradicting
  the derived 18+ stat on the same screen - removed at the frontmatter source,
  which also fixes the LIVE page's FAQPage JSON-LD.
- F6: the rail sat outside BOTH copy ratchets (marketing_copy_test globbed
  only themes/ layouts; RENDERED_GLOBS had no next/). Both extended.
- F4 a11y: FAQ section heading is a real h2 (was a styled p - 330 words
  invisible to heading navigation).
- F8: tenure derivation gets a rendered assertion, with its honest limit
  documented: derived == frozen until 2027, so it cannot discriminate this
  year and bites forever after.
- F7/F10/F11: exact-duplicate token aliased to var(--surface); measure back
  to ~68ch; comment corrected (the #fff-on-paper tonal step is 1.05:1 - the
  hairlines do the work).

Verifier's measurement worth headlining - the repo's own CSS evidence form:
next-pilot 3,105 B gzip vs single-service 21,586 B = minus 85.6% (minus 83.2%
netting out swiper). That is the number the clean-slate rail exists for.

Gates: bin/hugo-build clean; next_rail_test 4 runs 23 assertions green;
marketing_copy_test green with the widened globs; browser-verified at 1920 and
390, console clean. OKF entry rides this commit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>