feat(rows): five slots, and a window you are in is not one you have left

[lroolle]

The 5h strip is five cells again — but the future in them is empty, not
judged. v0.33.0 was right that 5h ▮▯××× was unreadable and wrong about
which half to cut. The × was the problem: a linear projection dressed as
ink, saying a third time what the badge (5h[38%@23:00]) and the notice
(5h caps ~14:20) already say with better gates and an exact time. The
hollow cells were never the problem — they were the axis. Ending the strip
at ▮ took the ruler away with the forecast, and left a bar that grew an
hour at a time and answered "how long have I got" with nothing.

So: five slots, always, one per hour, no dry cell in any of them.
5h ▃▃▮▯▯ is two whole hours after this one, read off the row without
arithmetic and without a second glance at the clock. Fixed width is the
other half of it — the row holds its shape for the life of the window
instead of reflowing every hour, which is the difference between an axis
and a bar that grows at you. An empty cell is a fact; a × is a guess, and
only one of those belongs in a ledger.

▮ now rides the real clock rather than the grid. five_period_start
rounds to five minutes so week_scan's cache key holds still across renders
— a resets_at that jitters by a second would re-run a whole-log jq pass
every render — and that rounding offsets every hour boundary by up to 2½
minutes. Invisible in a bar height; wrong exactly where this strip is read.
With the marker at 4 - floor(left / 1h) the hollow count is the whole hours
remaining to the second: at the 119-minute mark, three hours and one minute
left drew as two, and now does not.

N✕5h left no longer counts the window you are standing in. The row
draws it as ▮ and line 1 prices it as 5h[38%], so counting it again
made ▮ + 11 read as twelve, and the budget sentence beside it agreed with
the miscount. "Left" now means still to come: what remains after this
window closes, (7d left - 5h left), divided into windows — a stub at the
end of the week is still a window you can spend, so that rounds up.

The arithmetic has a property the old one did not: both clocks tick down
together, so the difference does not move. The count holds steady for the
life of a window and steps down by exactly one at each rollover. It was a
reading that drifted; it is a countdown now. windows_ahead is that
definition in one place, and the folded ...▯(✕N) prints what the budget
line computed rather than re-deriving it off a 34-cell grid that spans 170h
against a 168h period.

last window now means the week ends inside the one you are in — nothing
ahead of it, nothing to divide the surplus across. It used to fire at one
window ahead too, to skip a /win clause that would just restate the
headroom; calling two windows the last one to save a redundant clause is the
wrong trade. At one ahead the line keeps the grammar: 1✕5h left · 25.0%/win.

make install. The one-liner installed from GitHub and there was no
way to install the tree in front of you, so a working copy got there by
hand — and a stale hand-copy is how v0.28.0 once faked a red "7d dry" at
2%. make install runs the same install.sh with STATUSLINE_SRC set:
one installer, two entrypoints, no drift. It also refuses a
statusline.sh that does not parse — a broken statusline is not a worse
render, it is no statusline.

Three things the installer should have been doing all along, now on both
paths: it keeps the flags already on statusLine.command (rewriting
the whole command silently reverted --order and --debug on every
update), it writes through a temp file and renames (the script runs on
every render; a half-written one is a broken prompt), and it installs the
usage-insight skill beside it (STATUSLINE_SKILL=0 opts out).

make status reports installed-vs-tree drift, settings command and skill
state; make check is shellcheck + bats; make install-check gates the
install on both. make help lists the rest.

435 tests (was 427).