fix(widths): guarantee minimum column width of 1 (empty-column alignment)

[SayreBlades]

Summary

markdown-table-wrap-distribute-widths had two code paths with different minimum-width contracts. The shrinking path (table wider than available) already floored every column at 1 via (max 1 (floor ...)), but the fits path (the common case) returned raw natural widths — which are 0 for a column whose header and every cell are empty. Returning 0 misaligns the rendered table: the separator line draws a dash (via its own defensive max 1) where the row draws nothing, so every line after the separator drifts by one character per empty column.

This applies (max 1 ...) in the fits path too, so both paths agree that column widths are always ≥ 1.

Repro

A table with an all-empty middle column, rendered with markdown-table-wrap at a wide terminal:

Before (rows are 1 char narrower than the separator per empty column):

| field | density |  | field    | density |
| ----- | ------: | - | -------- | ------: |     ← separator is 1 wider here
| pos   |   0.734 |  | rotation |   0.012 |     ← rows drift right is broken

After (all lines aligned):

| field | density |   | field    |   density |
| ----- | ------: | - | -------- | --------: |
| pos   |   0.734 |   | rotation |     0.012 |

Why this is the right layer

The inconsistency is internal to distribute-widths: the function already believes in a min-1 contract (the shrinking path enforces it), it just didn't apply it in the fits path. Every renderer (markdown-table-wrap itself, and downstream consumers like pi-coding-agent's chat and the table-pretty overlay library) has to defensively (max 1 w) on the separator but not the row (or vice versa), and they drift. Fixing the allocation function once fixes all renderers, and the separator's existing (max 1 w) is no longer patching over a 0 that shouldn't have been returned.

The existing test markdown-table-wrap-test-widths-fit-naturally already asserted (should (>= (nth N widths) 1)) — the suite already encoded the min-1 contract; this just makes the fits path actually honor it for empty columns.

Scope

• One change in markdown-table-wrap-distribute-widths (fits path): wrap the returned natural widths in (mapcar (lambda (w) (max 1 w)) ...).
• No change to the shrinking path (already min-1), no change to compute-table-metrics (natural-widths remain a pure measurement of max visible width).
• No change to the budget math (border-overhead, total-natural) — those use raw natural widths for the fits/shrink threshold, which is unaffected; the min-1 floor only applies to the returned widths.

Behavior risk

None. A 0-width column cannot render at 0 anyway (it needs at least | |); bumping to 1 is strictly more correct. The shrinking path is untouched (it already capped at natural and floored at 1). The sqrt weights already used (max 1.0 ...) so empty columns already participated with weight 1.0.

Tests

Two new ERT tests in test/markdown-table-wrap-test.el:

1. markdown-table-wrap-test-widths-empty-column-min-1 — an all-empty column computes to width 1 (not 0).
2. markdown-table-wrap-test-empty-column-row-separator-align — a rendered table with an all-empty column has rows and separator of equal width (no drift).

Full suite: 156 tests pass (154 existing + 2 new), 0 regressions. Byte-compile clean.