feat: masonry board and fact grids, full-bleed width, contradicted hidden by default

[msitarzewski]

What

The Edge Board moves from a capped 1320px column with an equal-height auto-fit grid to a full-bleed masonry layout — at both levels, the board and the fact grids inside each card. Contradicted events come off the board by default. Plus one real bug found during verification.

Why

Three problems on the live board:

1. The 1320px cap wasted horizontal space on wide displays.
2. auto-fit stretched every card in a row to the tallest one, so a row containing one long event left ragged whitespace under its short neighbours. The fact grids inside each card had the identical problem — a one-line price sat in a box padded out to match a three-line yield note beside it.
3. Contradicted events competed for attention with breaking ones, which inverts the point of the board — the stream being ahead of the wire is the signal; a refuted claim is not.

How

• main is full-bleed with fluid gutters (clamp(18px,4vw,60px)) instead of a fixed column.
• Both grids are true masonry: grid-auto-rows:1px plus a per-child row span. Row spacing is the child margin-bottom, not row-gap — a row-gap is charged per 1px row and would coarsen the span quantization.
• The packing maths is shared as packGrid() rather than duplicated. Fact grids pack before cards are measured, because reclaiming their dead space shortens the card; measuring first would bake in the taller pre-pack height.
• All #feed mutations go through setFeed(), which assigns the row span. A node inserted without one collapses to 1px and paints over the footer, so this is load-bearing rather than tidiness.
• DEFAULT_HIDDEN keeps contradicted off the board unless its pill is selected. statusCount still counts the full item list, so the pill shows the true number — hidden never means gone.
• Card minimum drops 440px to 360px, giving masonry more columns to pack into.

Bug found while verifying

scroll-margin-top was a hardcoded 164px against a 169px header, so clicking a ticker headline landed the target card 5px behind the header. A constant can't be correct here — the operator telemetry rows are hidden from public visitors, so the header height differs per viewer, and it wraps taller as the viewport narrows (319px at mobile width, where the old constant would have hidden the card by 155px). The offset now derives from the measured header via a --header-h custom property, republished on resize and after each render.

Testing

Verified in a browser against 100 real events, at 1920px (4 columns) and 492px (1 column):

• 0 card overlaps, 0 fact-box overlaps, 0 horizontal overflow at both widths; footer clears the feed.
• All 100 fact grids packed; 0px of stretched box height remaining (was the reported symptom).
• Targeted cards clear the header by 14px at both widths, with :target active.
• Contradicted: pill reads the true count (2), board and ticker both exclude them, selecting the pill reveals exactly those 2, deselecting restores.
• Offline suite: no new failures. The 13 pre-existing failures (12 test_service_setup, systemd-dependent; 1 dashboard concurrency test) reproduce identically on main with these changes stashed — macOS-local. CI is the gate.

🤖 Generated with Claude Code