smooth leaves by sample support, not cube root

[andy-k]

Summary

When a leave has too few samples to trust its mean, the generator borrows a value from its
one-tile-swap neighbors. The old rule smoothed a leave when its draw-weighted count fell
below the cube root of the total -- a weight that counts how often the leave can be drawn,
not how many racks were actually sampled. It smoothed about 80% of leaves even when every
one was well covered, replacing measured values with borrowed ones for no reason.

Gate on the raw sample support instead: the total sampled count of the racks feeding a
leave, against a floor (WOLGES_GENERATE_SMOOTH_MIN, default 50).

Changes

• Smooth on raw sample support rather than the cube root of the draw-weighted total.
• Report how many leaves were smoothed and what share of the table that is, so a run that
  smooths nearly everything says so instead of doing it silently.
• Document the new knob, and re-measure the section's figures against the rule that ships.

Correctness

The old "smoothed and unsmoothed play the same" figures were measured under the cube-root
rule and describe a table nobody builds now. Re-measured across the game count, each table
against its unsmoothed twin over a million game pairs:

games        leaves smoothed   smoothed side won
50,000            89%               51.61   certain
200,000           76%               50.32   certain
1,000,000         52%               50.03   not resolved
5,000,000         26%               50.00   dead level
20,000,000         0%               identical file

Every row at or above 50, none below.

Test plan

• 20,000,000-game summary: english-generate and english-generate-no-smooth produce a
  byte-identical klv2 (nothing is thin, so nothing is smoothed).