fix(usage): locate claude in ~/.local/bin; drop stale codex tiers

[StuBehan]

Two Usage-tab fixes.

Claude usage doesn't show (regression from #130)

#130 moved Claude quota onto the claude CLI, resolved via ProcessOutput.claude(). That list only probed /opt/homebrew/bin, /usr/local/bin, and ~/.claude/local. Claude Code's current native installer symlinks the binary into ~/.local/bin, which isn't on a launchd-spawned app's minimal PATH, so the probe returns nil → hard-fail → nav.quota stays nil → .claude never enters availableUsageClients and the tab shows no Claude row (and no error, since other clients are present).

Fix: add ~/.local/bin/claude (first) to the probe list. Parsing of claude --print --output-format json /usage is unchanged and confirmed still matches the current output.

Codex shows stale windows with past reset times (pre-existing)

CodexQuotaProbe reads the newest ~/.codex/sessions rollout and renders its rate_limits with no freshness check. When Codex hasn't run recently, the newest rollout is older than its own window, so it shows old used_percent with a reset time in the past (e.g. "resets 1 week ago" for the 5h window).

Fix: in tier(), drop a tier whose resets_at is already in the past — that window has rolled over and the captured percentage is stale. When both tiers are stale, .codex drops out of availableUsageClients (same "no current data" behaviour as an unavailable Claude).

Testing

swiftc -typecheck panel/*.swift shared/*.swift clean. Verified the live claude /usage output parses and reproduced the stale-Codex case against a 9-day-old rollout locally.