Telemetry history: per-core CPU and temperature series on the machine board (#1789 follow-up) #2141

The machine board now keeps a history of CPU use and temperature for each core, in the same 15-second ring (machine_sample) as the rest of the machine source.

What is sampled (src/boards/sampler.ts, new src/boards/thermal.ts)

• CPU per core: each logical CPU is its own target, cpu:<nnn>, zero-padded so the CPUs list in order. Its value is the CPU's cpuN line in /proc/stat turned into a rate (busy ticks ÷ 100 Hz ÷ seconds), so at most 1 core.
• Temperature per core: each cpu:<nnn> also carries the temperature of the physical core it runs on, in °C. That comes from the coretemp chip's Core <K> input under /sys/class/hwmon, matched using the CPU's topology/physical_package_id and core_id. Two hyper-threads on one core therefore read the same thermometer.
• Other thermometers: each becomes its own series, °C only: package:<id> for a CPU package, sensor:<chip>/<label> for anything else (k10temp, acpitz, nvme…).
• Not guessed:

- A coretemp chip with no Package id label is not assigned to a package; its inputs stay sensors. - A reading outside −100…250 °C (what a broken sensor reports, e.g. −128000) is dropped. - Chip names and labels are cut to a short, plain set of characters before they become part of a target name.

• gaps() names what this machine can't read: no per-CPU lines, no temperature at all, or temperatures but none that map to a core.

Storage and query

• An additive migration adds celsius and celsius_max columns. The 5-minute roll-up keeps the mean and the max, like CPU and memory.
• The machine source documents both columns, with units.
• Board queries get a new filter op, starts (src/boards/query.ts), so a board can pick just the cores with target starts cpu:. It uses instr(), so % and _ match literally, and it is refused on number columns. docs/APPS.md describes it.
• The "What this machine measures" card on /boards/data?source=machine links four ready-made queries:

- CPU per core, last hour - max °C per core, last hour - a 12-hour heat strip of CPU per core - every thermometer, last hour

Tests

• New test/boards-thermal.test.ts builds a fake /sys with two packages of hyper-threaded cores, and its expected rates and °C are worked by hand. It covers the per-core mapping, broken readings, duplicate sensor names, roll-up of °C and the starts op.
• test/boards-page.test.ts follows the card's links and checks they show only the cores.
• I changed two existing tests in test/boards-sampler.test.ts because the data grew, not because they were wrong:

- the fixture's paths gained sys; - the expected gap count went from 5 to 7, and the startSampler rows now include cpu:000 and package:0.

• bun x tsc --noEmit is clean. The boards, schema-additive and five cross test files pass (266 tests).
• On this 80-CPU node all 80 CPUs got a temperature, plus 2 packages and 4 other sensors, with no gaps; the first (cold) read took about 20 ms.

Choices to look at

• Disk: each CPU is a full ring target, so an 80-CPU node can hold up to about 1.2 M more rows (around 80 MB). That is noted in docs/APPS.md.
• Sampling rate: samples stay at 15 s rather than the 2-second live rate in #1789.
• Row cap: the 12-hour heat strip fits the 1000-row cap only up to 83 CPUs; on a bigger node the oldest hours are cut off.
• Not done: the node's node target gets no temperature of its own, and there are no AMD per-core temperatures, because k10temp reports per die, not per core; those go out as sensor: series.

🤖 Generated with Claude Code

Closes #2081