feat(terminal): read COLUMNS from Claude Code before probing

[lancemarks]

Summary

Claude Code 2.1.153+ sets COLUMNS and LINES to the current terminal size when it runs the status line (changelog, docs: "Sizing output to the terminal"). getTerminalWidth() now reads COLUMNS right after the CCSTATUSLINE_WIDTH override, so on current Claude Code the ancestor walk never runs.

Motivation

• Perf. On macOS/BSD the walk still spawns ps twice per ancestor plus stty on every refresh. Linux already avoids that via the /proc probe from perf: stop re-probing the terminal width on every line of every render #501. Measured on an M-series Mac (built dist, 40 interleaved runs, minimal terminal-width + model config, same process depth as a npx ccstatusline status line):

  	median	p10	p90
  walk (no COLUMNS)	193 ms	176 ms	237 ms
  COLUMNS=120	136 ms	127 ms	162 ms

  About 57 ms saved per render.

• Robustness. It closes the "no ancestor owns a TTY" gap from Terminal width probe fails under Claude Code ≥ 2.1.139 (no controlling TTY for spawned statusline) #376/feat(terminal): honor CCSTATUSLINE_WIDTH env var to override probe #380 without the user setting anything, and gives Windows a width source by default, since probing is disabled there.

Change

src/utils/terminal.ts: after the override block, a positive-integer COLUMNS is memoized and returned. Order is now:

1. CCSTATUSLINE_WIDTH (explicit user override still wins)
2. COLUMNS
3. session "no TTY" cache
4. native /proc probe → ps/stty walk → tput cols

COLUMNS goes before the cache so that a stale cached "no TTY" result can't hide a real width. Missing, empty, 0, or non-numeric values fall through to the existing path, the same way #380 handles bad input.

LINES is deliberately not read, since nothing in ccstatusline uses terminal height yet.

Test plan

Six new cases in src/utils/__tests__/terminal.test.ts (COLUMNS is also cleared in setup/teardown so an ambient value can't leak into the probe tests):

• COLUMNS short-circuits probing (no execFileSync calls)
• CCSTATUSLINE_WIDTH beats COLUMNS
• COLUMNS=0 / COLUMNS=wide fall back to probing
• COLUMNS applies on Windows
• COLUMNS wins over a cached session "no TTY" without touching the cache

$ bun test
 2364 pass
 0 fail
$ bun run lint   # clean

The three behavior cases fail against main, and the fallback and precedence cases pass on both.

Risk

• If someone exports COLUMNS in their shell rc and runs Claude Code older than 2.1.153, the status line inherits that fixed value where the probe used to find the live width. This is the same trust tput cols (the existing last fallback) already gives COLUMNS through ncurses, and CCSTATUSLINE_WIDTH still overrides it.
• The TUI reads process.stdout.columns and never calls this path, so it's unchanged.