fix(proposal): move the spacing onto the layout boundary and make the type fluid

[scheemunai]

The boundary, and I caused it

proposal's 34rem block toggled type scale and spacing while declaring nothing structural — a reader between 545 and 599 received a measure set for a layout they did not have.

That was #29. It lifted .hero{grid-template-columns} out of the 34rem block into its own 37.5rem block to fix a horizontal overflow, and left the old boundary purely typographic. Verified against history:

1fa8031  34rem block: .hero{grid-template-columns:minmax(0,1fr)}   1 structural declaration
a681dee  (#29)                                                     0 structural declarations

The fix

Rather than give the old boundary a structural rule it does not need:

• the spacing moves to 37.5rem, where the hero actually restructures — type, spacing and layout now move together, which is what the correctly-classified boundaries already do;
• the two stepped type sizes become clamp() — .mast .lead and .move h2 — so there is no step to misplace.

That is the shape runbook and changelog already use: one breakpoint, fluid type.

media blocks              2 -> 1
structural decls in it    1
font-size in any block    0
step at 540..550          none
remaining boundary        600->601, +541, and it restructures

Verification

390     2,687 -> 2,687   unchanged
768     1,995 -> 1,987   8px better
1,440   2,060 -> 2,060   unchanged
overflow  0 across 992 widths — 320-700 step 1, 700-1920 step 2

window.innerWidth is asserted before every recorded measurement; the harness refuses rather than records on mismatch.

Nothing removed. No line-height touched. pnpm check clean; pnpm exec vitest run — 151 files, 1,096 tests, all passing.

A note on the test run

My first full suite showed 3 failures against a baseline of 2, and I nearly reported that my change added one. The named test passes in isolation three times in both trees, and a second full run of the fix tree is completely green — 0 failures. Both earlier runs were flake. I had compared two single samples of a non-deterministic measurement and read the difference as signal.

🤖 Generated with Claude Code