feat(tui): redraw the small cross as a three-row sign with sub-cell fillets

[plombeer31]

Stacked on #227 → #226 → #225 → #224 → #223 → #222 → #220.

The mark

   ▗█░
 █████░
   █▘░

Three rows, one-cell arms, a quadrant block tucked into each concave corner, keeping the one-column bevel.

Why

The small mark was the five-row bevelled cross. At the two places it actually appears — the rail lockup and the setup headers — five rows of logo out-shout the two lines of text beside them. It is a sign at this size, not a reproduction.

The fillets. The concave diagonal (top-left, bottom-right) is what distinguishes this mark from a plain cross, and at one cell per arm there is no room to draw it in whole cells. A quadrant block puts the ink in the corner it belongs to at half the size — the only sub-cell tool a terminal offers. The hard 90° corners stay empty: filleting all four would make the mark 4-fold symmetric, which is a different logo.

The part worth reviewing

This size is constructed, not rasterised. Every other mark samples the bezier path from assets/logo.svg. This one cannot: the arm is a quarter of the box, so a one-column arm implies a five-row box at a 2.2:1 cell, and fullGrid(5) rounds straight back up to a two-column arm. The proportion (arm 1, bar 4×arm + 1 for centring) is written out in renderSmall instead, with the reasoning in the doc comment. It is still generated code — logo-art.generated.test.ts runs --check and fails on drift — but it is derived from the rule rather than from the path, and that deserves a reviewer's eye.

ASCII keeps plain cells (no quadrant glyphs in that stroke); the charset test still passes.

Knock-on effects

The mark lost two rows, and things measured it:

• SIDEBAR_CHROME_ROWS 16 → 14, so the rail's sessions/tasks budget grows by two rows (24-row terminal: 4/2 → 6/2).
• LOGO_METRICS.mini 9×5 → 6×3, which gives the splash one more tip row on a small window.
• RailBrand seated its two text rows under two blanks to centre them on a five-row mark; one blank now.

All three are pinned by existing tests, which is how each of them surfaced.

Verified

Captured at 100×30: the setup header and the agent's rail lockup both draw the new mark, with atomic-agent / v0.3.6 seated on the bar row.

Full suite: 5205 passed. fs-glob-real and send-message-concurrency fail on main too; llm-health-poller is the usual parallel-load flake and passes in isolation.