Totals over the window, a readable time axis, and a build stamp

[evandhoffman]

Closes #35.

Three related pieces of one session, in one pull request because they were only
ever mergeable in order — see the commits for the separable history.

1. Totals over the window

Every legend entry gains a second number: how much moved over the window the
chart draws, beside how fast it is moving now. Network and disk in bytes,
packets and disk operations as counts.

Disk                     5 s
                      Totals
● Read   12.8 MB/s     35 MB
● Write  13.0 MB/s     14 MB

The sum is already in the buffer. A rate sample is the mean over the gap
before it, so Σ(rate × preceding gap) telescopes back to exactly the counter
delta. No counter history, no source changes, nothing read that the ring buffer
does not already hold — so it stays well inside the rule that v1 writes nothing
to disk. The round-trip test is the premise: differentiate a synthetic counter
with RateTracker, integrate it back, expect the original delta.

Three things it refuses to fake:

• It reports the span it covered. Ten seconds after launch the buffer holds
  ten seconds, and "2 min" over a twelfth of that looks exactly like the right
  answer — which is why nobody would check it.
• A gap nobody sampled is not traffic. An interval wider than maximumGap
  is clipped, so a laptop back from sleep cannot credit one sample with an hour
  that never crossed the wire. AppModel.totalGap follows the Sampling tab
  rather than assuming 0.5 s.
• Fewer than two samples has no total — nil, not zero, the same answer
  RateTracker gives on a first read.

Network totals in bytes while the rate stays Mbit/s: a link is quoted in
bits, a volume in bytes. MetricUnit.accumulation holds the ÷8 as its only
definition, derived from the unit like direction and composition.

A card with totals draws a Grid — headed, right-justified — rather than the
wrapping FlowLayout. A magnitude is read by where its last digit sits, and
wrapped entries put them at four different left edges.

2. A time axis that does not collide

From a screenshot: 08:52:3008:53:0008:53:30. Three things were wrong at once.

• The room was measured off the chart, not the plot. A flat 50-point
  allowance stood in for the y-axis — right for one card, generous for every
  card whose numbers are long, which are the cards whose plots are narrowest.
  Now read from the real plot rect via chartBackground. Not chartOverlay:
  an overlay would swallow the mouse-down a tile's drag needs.
• Every stride was costed at the same label width, though a stride of a
  minute or more shows no seconds. Measured: 20 points against 32.
• The labels were wider than they needed to be — own size now, and the hour
  loses its leading zero.

Rule 2 now beats rule 1. "Never fewer than two labels" used to win on the
grounds that slightly tight beats bare. Right about slightly tight, wrong about
the output: four labels 32 points wide with centres 40 apart. Two labels are
still what the walk reaches for; what is gone is the promise.

Turn the time labels sideways, new in the Charts tab, off by default: on its
side a label costs its line height, so the narrowest card fits four times where
upright it fits two.

3. A build stamp in the title bar

Monitor
v1.4.0-11-g0ac6bafa · Aug 28 09:56

There to identify which build produced which screenshot. White on black in the
system font — deliberately not the panel's palette or the cards' monospaced
face. It is the one thing in the window that is not a reading: everything else
is a measurement styled to be scanned, and this is a label on the photograph.

Two facts, two sources. The commit is stamped at build time by a new
StampCommit prebuild plugin (git describe --tags --always --dirty) — a
running program has no other way to know, and a checked-in constant is one
somebody has to remember to update. The build time is read at runtime from the
executable's modification date: stamping it too would rewrite a source file on
every build and recompile MonitorCore every time, for a fact the filesystem
already has. The plugin rewrites its output only when the hash changes, for the
same reason.

Three things this turned up on the way:

• macOS 26 wraps toolbar items in a shared glass capsule, which made the stamp
  dark-on-light and clipped it. It opts out with
  sharedBackgroundVisibility(.hidden), gated to 26 and additive.
• The window's own title is removed, or the name draws twice. toolbar(removing:)
  is macOS 14 but the .title kind is 15, so 14 gets an empty navigationTitle.
• Removing the title cost the toolbar its trailing alignment — the title had been
  taking the slack in the middle, not the items' placement. They are
  .primaryAction now, which says what the layout actually depends on.

Verification

215 tests in 26 suites, swift build -c release, and swiftformat Sources Tests Plugins --lint all clean. Lint now covers Plugins; ci.yml updated to match.

ChartAxisTests is rewritten against the new contract, keeping every existing
claim and adding the ones this needed — chiefly labels never overlap, swept
across every window, six plot widths, both orientations and forty phases each.

All three pieces were looked at in the running app and revised from what they
actually looked like, which is where most of the above came from.

https://claude.ai/code/session_01EqTTtmt4fyNtMxjc2ZVBzj