perf: stop re-probing the terminal width on every line of every render - #501

Persists probed widths to ~/.cache/ccstatusline/terminal-width.json, keyed
by session_id (width is per-terminal, so a global value would be wrong).
Atomic tmp+rename writes, entries pruned after an hour, corrupt/unwritable
cache treated as a miss and never fatal.

readCachedWidth returns a wrapper ({width} | null) so a cached null -- 'we
probed, there is no TTY' -- is a HIT, not a miss. Collapsing that would make
the no-TTY case re-probe forever, which is the bug this exists to fix.

ttlSeconds 0 disables caching (always probe). Note this deliberately differs
from gitCacheTtlSeconds, where 0 means 'never expire'.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Claude Code spawns the statusline without a TTY, so probeTerminalWidth()
returns null. Callers read it as 'context.terminalWidth ?? getTerminalWidth()',
so the null re-triggered the full ps/stty ancestor walk once per configured
line. Cache the probe behind an explicit hasProbed flag so the negative
result is memoized too.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

execSync with shell:'/bin/sh' spawned a shell in addition to the command,
doubling every probe; the stty|awk pipe cost three processes. Call the
binaries directly and parse 'rows cols' in JS. Drops the legacy
'stty size < /dev/tty' redirect form, which required a shell and is
redundant with the -F/-f forms.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Chain is now: CCSTATUSLINE_WIDTH override -> in-process memo -> shared L2
disk cache -> probe (native /proc+TIOCGWINSZ on Linux, ps/stty/tput
elsewhere). Only the entry point supplies a session_id, so only it reads or
writes the shared cache; renderer.ts and widgets/TerminalWidth.ts keep their
zero-arg calls and hit the memo.

Adds terminalWidthCacheTtlSeconds (default 5, 0 disables) to Settings and
RenderContext, and backfills it in two widget test fixtures that build a
full Settings literal (zod .default() makes the key required on the output
type).

Reads process.platform at call time in terminal-native's default deps: a
module-load snapshot ignored the per-test platform pin and ran the real
Linux probe under darwin-pinned tests.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Co-Authored-By: Claude Opus 4.8 <[email protected]>

- Serialize writes to the persistent width cache with a best-effort
  file lock (O_CREAT|O_EXCL, stale-lock recovery), closing a
  read-modify-write race where two concurrent ccstatusline processes
  (e.g. two Claude Code sessions) could silently drop each other's
  cache entry.
- Only ever persist a "no TTY" probe result across process boundaries;
  a discovered numeric width is no longer written to the cross-process
  cache, since it's keyed by session_id and the same session can be
  resumed in a differently-sized terminal, which would otherwise serve
  a stale width for up to terminalWidthCacheTtlSeconds.
- Correct the README's claim that the cache is "shared ... across
  sessions" -- it's deliberately isolated per session_id, which is the
  point of keying it that way.
- Remove context.terminalWidthCacheTtlSeconds: it was write-only dead
  state (the TTL is already consumed by the getTerminalWidth() call
  that builds the context).
- Add a Terminal Width Cache TTL entry to the TUI's status-line config
  menu, mirroring the existing Git Cache TTL entry -- the setting was
  previously only reachable by hand-editing settings.json.
- Add integration/wiring tests for the sessionId/ttlSeconds path
  (previously only unit-tested in isolation) and for the new locking
  behavior.

…ocks

Around twenty suites replace child_process wholesale with
vi.mock('child_process', ...). That factory form is not file-scoped under
the bun runner, so the replacement stays installed in the module registry
for every file that runs afterwards.

When one of those suites runs before this one, vi.spyOn hands back the
already-installed mock together with its accumulated call history instead
of a fresh spy. The stderr-silencing test iterates every recorded call and
asserts stdio is ['ignore', 'pipe', 'ignore'], so it was inspecting git and
terminal probes from other files, which legitimately use
['pipe', 'pipe', 'ignore'], and failed.

The test passes in isolation and fails in the full suite, which is why CI
reported it as the only failure. Clearing the call history when the spy is
installed scopes every assertion in this file to the calls its own test
made.

Co-Authored-By: Claude Opus 4.8 <[email protected]>