chameleon 4.4.45

Fixed

• TEST_PAIRING_MIN_SAMPLE was left at 10 when the sibling convention floors
  were measured down to 5 (v4.4.23), so test-pairing derived for almost nothing
  even after v4.4.43 taught it the flat tests/ layout. The two thresholds
  answer the same question -- how many siblings make a convention trustworthy --
  and conventions.py says outright that they "should not disagree". 10 sits
  ABOVE the natural cohort size of a real repo: across the fixture matrix only
  1 of 7 archetypes (ts-plain), 1 of 12 (py-django) and 1 of 10 (rb-plain) clear
  it, so a fully-paired repo still produced test_pairing: {} and the stale-test
  advisory riding on it stayed silent. Lowered to 5 to match MIN_SAMPLE_SIZE.

  test_below_sample_floor_returns_empty was sized at a literal 9 (chosen when
  the floor was 10) and would have silently stopped testing anything; it now
  sizes off MIN_SAMPLE_SIZE itself so the floor and its guard cannot drift
  apart again.