actop 1.9.8

Skip 1.9.1 through 1.9.7 — install this one instead. Those seven were cut in
rapid succession (six of them inside forty minutes) while the release pipeline
itself was being stabilized: each fixed the flow that had just shipped the one
before it, and none carried a user-visible change worth a release of its own.
1.9.8 supersedes all of them and is the first release in that line with a real
fix for users. CLAUDE.md now gates releases on user-visible change so this
does not recur; the intermediate tags stay published because PyPI artifacts
already point at them.

Fixed

• Narrow grid columns no longer clip the tail off metric rows. At
  _GRID_MIN_WIDTH (96 cols) each box gets ~44 content columns, and the longest
  rows are wider than that: the P-CPU summary lost the % off
  avg 41% · max 41, and every residency row lost high37 outright (that one
  overflowed up to ~110 cols). The rows are now width-adaptive instead of
  hard-clipped — carried open since 1.5.0.
  • Rolling context degrades in three tiers, each row choosing the widest that
    fits its own measured width: avg N% · max N% → ⌀N ▲N% → ▲N% →
    nothing. Applies to every labelled row (P-CPU/E-CPU, GPU, ANE, RAM, Mem BW,
    the four I/O rates, CPU/GPU/Package power). The reading itself is never
    shortened to make room for its annotation.
  • Residency rows size their density bar to what the breakdown leaves (16 chars
    down to 8, dropped below that), so the four percentages always survive. The
    bar is budgeted against the widest the breakdown can ever be, so it does not
    breathe a column every time a percentage crosses 10.
  • The fit is against each row's high-water widths, not the current frame's, so
    a row on a tier boundary holds its form instead of flipping between the
    spelled-out and compact shapes as digits come and go (avg 8% → avg 10%,
    987MHz → 1987MHz).
  • ? help and docs/SPEC-system.md §6.3 document the ⌀/▲ compact form.