v2.4.0 — Topic-level tiering in the compact index

Closes a gap an audit of the brain's own claims turned up: the README promises the per-session
cost stays bounded "not by how much history you keep" — true across projects (HOT/WARM/COLD
tiers, since 2.0), but not true within a single project, whose full topic list rendered
unconditionally no matter how large it grew.

Backward compatible, still zero required runtime dependencies (Python 3 stdlib only).

Added

• Topic-level tiering in the compact index. Same idea as HOT/WARM/COLD, one level down:
  ⚠ in-progress topics are always shown in full, however many there are; once a project passes
  15 topics, its finished ones (✓v/✓d/✗/⨯) start collapsing — the most-recently-dated
  ones stay expanded up to that cap, the rest fold into one +K more, oldest <date> line. Pure
  function of index.md, deterministic, no new file format.
• brain-check's existing "project has >15 topics" warning now references the same threshold
  (_compact.TOPIC_THRESHOLD) instead of a second, independently-defined 15, and its wording
  reflects that the compact self-corrects now even though the underlying files still benefit from
  consolidating.

Full details in the CHANGELOG.