Waterfill column-width distribution - #3

Replace the ad-hoc proportional allocator with an analytical waterfill: find the unique level L where sum(min(nat, L·√nat)) = budget, then round via Webster/Sainte-Laguë.

The old algorithm had binary thresholds for small-column protection and tight-budget fallback that caused columns to shrink when the terminal got wider. The waterfill is monotonic by construction — no column ever loses width when more space becomes available.

The cliff

A table with columns | ID | Name | Category | Description | (natural widths 2, 12, 11, 93) hits a catastrophic discontinuity at width 39. The old algorithm suddenly protects Name and Category at full width, starving Description to 1 character — one letter per line:

Before (w=38 → w=39):

w=38: ID=1  Name=4   Category=4   Description=16
w=39: ID=2  Name=12  Category=11  Description=1   ← one letter per line
w=40: ID=2  Name=12  Category=11  Description=2

After — smooth progression, no column ever shrinks:

w=38: ID=2  Name=5  Category=4  Description=14
w=39: ID=2  Name=5  Category=5  Description=14
w=40: ID=2  Name=5  Category=5  Description=15

Benchmark

resize-all improves ~15% (fewer GC cycles from dropping the per-cell longest-word scan):

            main        waterfill
resize-all  ~1435ms     ~1210ms
GCs         268         220

Changes

• Replace four-phase distribute-widths with analytical waterfill + Webster rounding
• Remove min-word-widths from compute-table-metrics (no longer used)
• Remove --longest-word-width helper and --small-column-threshold constant
• Net −31 lines of production code