Count memory the way Activity Monitor does

[jondkinney]

The bug

The memory gauge counts every page that is not free or speculative as used, which sweeps in the file cache. On a 128 GB machine it reads 116 GB / 91% while Activity Monitor and iStat Menus both report 63%:

	EdgeControl 2.3.1	Activity Monitor / iStat
Memory used	116.0 GB — 91%	~83 GB — 63%
Pressure	91%	7%

The error scales with how much RAM the machine has spare to cache with, so it's worst on exactly the machines this dashboard is aimed at.

Memory pressure has a second problem: it's derived from the same free-page count, so it is effectively a duplicate of the used percentage. Both gauges move together and sit pinned near the top, which means a genuine pressure spike — the thing that gauge exists to show — has nowhere left to travel.

The fix

SystemMetricsService.currentMemorySnapshot():

• Used = app memory (internal_page_count − purgeable_count) + wired + compressed, the three figures Activity Monitor adds up. File-backed and purgeable pages are cache the kernel hands back on demand, so they aren't counted.
• Pressure = (wired + compressed) / physical — only what the kernel cannot reclaim, which is what memory_pressure reports as unavailable. On the sample above that gives 5% against iStat's 7%.

Incidentally this also drops a small double-count: vm_statistics64.free_count already includes speculative pages, so the old free_count + speculative_count counted them twice.

Tests

The arithmetic moves into a MemoryReading struct so it can be exercised without reading the live host. Tests/EdgeControlTests/MemoryReadingTests.swift runs the vm_stat page counts from that 128 GB machine through it and pins used and pressure to what Activity Monitor showed, plus the edge cases (clamping, purgeable_count > internal_page_count, zero physical size). Reverting the source to the old formula fails them with exactly the 116.1 GB / 90.7% from the screenshot.

Behaviour is unchanged otherwise: same struct fields, same call site, same fallback to the previous sample when host_statistics64 fails.

One note

I could not run the suite on a clean main — on Xcode 16.4 / Swift 6, main fails to compile before any of this (PluginWidgetProvider.swift:58, non-Sendable NSImage in a Sendable struct, and main-actor isolation errors in LayoutEngineTests). I verified these tests with those build errors patched locally. Happy to send that as a separate PR if it would be useful.