perf(sysstat): read the focused pane's process tree from /proc, not ps

[tfyl]

Why

While a pane is focused, the board refreshes its process line (CPU, RAM, what the pane runs) every 1.2 s, and each refresh forks ps twice. The first fork lists every process on the machine to report on the few under one pane. On a 1,636-process host that costs 159 ms per refresh, about 13% of a core spent on ps alone. A 60 s execsnoop of a live board counted 80 ps forks a minute.

Summary

• The new sysstat.Tree(pid) walks the pane's tree from its root through /proc/<pid>/task/*/children, using the walker the poll pass's TreeSampler already relies on. It forks nothing.
• PCPU is computed the way ps computes %cpu: each process's CPU time over its lifetime, using starttime from /proc/<pid>/stat and /proc/uptime. ScaleToHost then gets the same input it got from ps.
• Without /proc child lists (macOS, or a kernel without CONFIG_PROC_CHILDREN), Tree falls back to Trees unchanged.
• The two focus-capture call sites use Tree. Trees, and the TreeSampler seeding scan that tests check against ps, stay as they are.
• The other ps sites are left alone because no timer drives them. The child sweep's background-work check runs only for a finished child about to be retired, and SessionOfProcess runs once per caller, not on a clock.

Numbers

BenchmarkFocusedPaneTree, 50 iterations on a host with 1,636 processes:

	per refresh	ps forks
Trees (before)	158.9 ms	2
Tree (after)	0.053 ms	0

Test plan

• TestTreeAgreesWithTreesWithoutForkingPs: on a fixture tree, Tree matches Trees exactly on process count and children, keeps RSS within TreeSampler's 2× bound, and keeps PCPU within 0.5. A ps shim on PATH proves it never forks; passes 5/5
• go test ./internal/sysstat/ and go test -short ./internal/ui/ pass, run with env -u TMUX -u TMUX_PANE and a scratch TMUX_TMPDIR
• Nothing the board draws changes (same fields, same scaling), so this PR has no capture

🤖 Generated with Claude Code