actop 1.4.15

Fixed

• Memory bandwidth reading: each DCS BW histogram bucket is now represented by
  its midpoint instead of its upper edge. The PMP / DCS BW /
  AMCC RD+WR state names are bucket upper edges in 32 GB/s steps, so the
  bottom bucket spans 0-32 GB/s but scored a flat 32. On an idle machine ~99%
  of the residency sits in that bucket (measured on an M4 Max: 4351 of 4396
  samples), so Mem BW was pinned at a constant 32.0 GB/s with a flat chart
  line and an avg/max barely above it — real idle traffic of a few GB/s was
  indistinguishable from 31 GB/s. Each bucket is now weighted by the mean of its
  own edge and the previous one (the first bucket's lower edge is 0), derived
  from consecutive edges rather than assuming a fixed step.

  Reported values shift down by half a bucket width (~16 GB/s per
  controller die). Measured on an M4 Max via Monitor.get_snapshot(): idle
  went from a constant 32.0 to 16.85 / 17.00 / 16.82 / 16.87 GB/s over
  consecutive samples — no longer pinned — and a 4-thread memcpy workload
  read 109.8 GB/s, so loaded readings still track proportionally.

  Note this does not give sub-bucket resolution: the bottom bucket is 32 GB/s
  wide and holds nearly all idle residency, so an idle machine now reads ~16
  GB/s (that bucket's midpoint) rather than its true few-GB/s traffic. The
  histogram simply carries no finer detail below 32 GB/s, so the idle chart
  stays relatively flat — just at a defensible value instead of the bucket's
  ceiling. Chart and MEM-BOUND alert scaling are unchanged; both still
  normalize against the session ratchet from 1.4.13.