The last of the audit list: the type scale wins, the card headings get names, and the reveals reach the keyboard

[datlechin]

Everything left from the seven-lens consistency audit that was still real when
re-checked against the current tree.

Ninety-four elements were arguing with the type scale

Eighty-seven of them put leading-relaxed on text-sm — the site's body copy
rung — so almost every 14px paragraph overrode the 1.55 the metrics curve
suggested.

The overrides were right and the scale was wrong. 14px prose wants more air
than a 14px table cell. --text-sm--line-height is 1.625 now and all
ninety-four are gone. A scale that every paragraph has to argue with is not a
scale.

Measured before deciding: 103 of the text-sm sites are prose, 43 are
single-line UI, where the difference between 1.55 and 1.625 does not render at
all. The seven on text-base and non-mono text-xs also go, because 1.6
against 1.625 says nothing. Two survive on mono code blocks, where the extra
leading is a real choice about reading code.

The card heading forked three-to-twenty-three with no rule

There genuinely are two roles: a heading that titles a cell which is the point
of that cell — the download panel, a pricing tier — and one that titles a single
entry in a list of like things, where a larger size would make every item shout.

Both are named now, PANEL_TITLE and ITEM_TITLE, sitting beside CELL_DENSITY
where the other cell decisions live. The inconsistency was never that two sizes
existed; it was that nothing recorded which was which.

Eight hover reveals had no keyboard equivalent

Every one sits inside a group that is a link, so a keyboard user tabbing to it
saw nothing change: the database tile's arrow, three sponsor marks, the engine
monogram, both blog list headings. They pair with group-focus-within now.

Two of those additions landed outside their className on the first attempt —
a regex matching across quote boundaries. Typecheck caught one, a per-line check
caught the other, and all three sponsor marks were re-verified to have kept the
parity after repair.

Duplication

The copy control existed twice, down to the same 1200ms timeout, the same
two toast strings and the same cleanup effect — with forked chrome, one having a
visible boundary and one not.

The closing CTA existed twice, differing in one sentence: the OS requirement
line, which is the one thing a reader checks before downloading, and the two
pages disagreed about it.

The benchmarks grid left a border hanging. Below sm its three cells stack and
each draws border-b, so the third drew one immediately above the section's own
rule. The other four hand-rolled grids were checked and are correct — they put
the border on the first of two cells, not on every cell.

Checked and deliberately left

• Container width="md" on four pages. It narrows the text to a readable
  measure, which is right for long prose, and does not affect the rules — those
  bleed past the container by design.
• CELL_DENSITY written as a literal at 25 sites. All twenty-five agree on
  the same value, so this is a maintainability wart, not an inconsistency, and
  converting twenty-five className="…" strings to template literals is exactly
  the mechanical churn that just put two class names outside their attribute.
  Not worth the risk for zero visible change.
• leading-relaxed on two mono code blocks. Deliberate.

Verification

typecheck, build and pint clean. 107 tests / 1853 assertions with
REQUIRE_SSR=1. All nine page types re-rendered: one <main>, one <h1> and
zero coincident rules on every one.