+ Pinned columns in the window, a refresh picker in both front-ends, and a column copy

[Hawkynt]

Six boxes in §11, §12, §57 and §94, and the two defects that looking at the
pictures turned up.

§11 — freeze and pin columns, in the window as well

The toolkit grew a horizontal offset and a Frozen flag on its columns, which
is the seam this row was refused for. The window keeps a count of leading
columns rather than a set of ticked ones, the same as the terminal: the list
draws its pinned run as the leading run and stops at the first column that is
not, so a pinned third column with two scrolling ones in front of it is a
layout nobody can draw. # moves the boundary in the terminal, Ctrl+Shift+P in
the window, and the count survives a restart on both sides.

Two defects came out of the first capture of it, neither of which any test
could have seen:

• The header hit-test added the scroll offset to every x, so a click on a
  pinned caption picked whichever column had slid underneath it and a resize
  started there dragged the wrong boundary.
• The list fills a row's ground once for the whole row and then draws the
  pinned columns in a second pass over the scrolled ones, so a pinned cell
  arrived with somebody else's value already in it and put its own on top.
  Every pinned cell in the first picture showed two values at once.

§12 — a refresh picker, and a difference between pausing and asking by hand

§12 has listed eight refresh settings from the beginning and neither front-end
could reach any of them: the terminal could pause, the window could not even do
that, and the rate was fixed for the life of the process. Both now have a
picker over one shared list of rates — View ▸ Refresh in the window, d in the
terminal, which is the letter every other terminal monitor binds a sampling
delay to.

Paused and by-hand are two entries. Both stop the tick; only the second is
remembered, because a monitor that opened paused because it was paused when it
was last closed is a monitor showing a table of nothing at all. The rate
underneath is kept either way. Neither disturbs the list — nothing is rebuilt
while the tick is off — and the terminal's loop reads the rate each time round,
so a rate chosen mid-wait shortens the wait that is running.

§11 — a column copy, and the full forensic set

A column copy was refused for want of a cell selection to take it from. There
is none in either front-end and none is needed: a column copy needs a column,
and both have had a column cursor since their headers grew gestures. Ctrl+Y in
the terminal, Ctrl+Shift+D in the window, over the ticked rows if any are
ticked and everything on screen otherwise.

The forensic set is the expert set plus the two halves it was missing — who a
process really is, and what it is doing to the disk. Deliberately the dearest
set in the file: the package, the digest and the descriptor count each cost a
reading the sampler does not otherwise take.

It is also twenty-five columns, which turned up the third defect. A set with
more columns than the terminal can hold even at their floor was reserving room
for the ones past the edge, and the columns that were drawn paid for it: the
process name came out six characters wide at a hundred and sixty columns, which
renders every row as kthrea. Room is kept only as far as the screen can
reach now; the rest is reached by scrolling sideways, which is what that scroll
is for. The name gets thirteen characters instead of six.

Verification

0 warnings, 1749 tests green, --self-test agrees with the runtime on all 37
checks. Frames captured at 80×24, 120×30 and 160×50, under a UTF-8 locale and
under LANG=C with TERM=dumb; the default frames are byte-identical to
before, and the picker fits inside eighty columns with nothing clipped. The
window was photographed with the forensic set pinned and scrolled to its far
end, which is the state a picture of the default six columns cannot show.

The benchmark's allocation gate reads 202 and 211 bytes a sample against a 512
floor. One earlier run read 724 against a 717 allowance while the capture
script's workload was still spawning processes — that allowance is
512 + 1024 × new processes per sample, and nothing in this branch is in the
sampling path.

Left open, and why

• §57.2 friendly/tree/flat — friendly needs §13's notion of an application, and
  there is nothing to read one off.
• §57.3 s as sort — a deliberate binding, documented as such, and keys.conf
  can swap it.
• §83 application, publisher, aggregations — need §70's signature verification
  and §82's sums.
• §83 package grouping switched on mid-session — reading a package costs a
  database lookup per image and the probe's expensive readings are chosen when
  it is built (§5.4). Making them switchable at run time is the probe's
  business, not the front-ends'.
• §12 dead processes visible for one cycle — the styling and the palette exist
  on both sides, but a ghost row needs a record in the snapshot after the
  process has gone, which is the sampler's to provide.

Two things noticed and deliberately not touched, both pre-existing and both in
files this change has no business in:

• The Working set header loses its W behind the CPU % header beside it at
  the widths this machine's settings file happens to hold. It is in the
  committed screenshot too, and it is the toolkit's header painter.
• Start time at its registry width cuts the last digit off a full timestamp —
  2026-08-21 22:03:2. Visible only in the sets that carry the column, which
  is why nothing had photographed it. The width is in FieldRegistry.