statusline: native context %, session-limit segment, wrapped + sorted MCPs

[ldayton]

A batch of statusline improvements that lean on the data Claude Code now hands the script on stdin instead of recomputing it.

• Context %: replace the transcript-tailing token sum with Claude Code's pre-calculated context_window.used_percentage. This drops a ~60-line file read per render and fixes an accuracy bug — the old calc folded output_tokens into "used", whereas the canonical percentage is input-only (input + cache_creation + cache_read), so our "% left" read hotter than /context. The 80 - x "% until auto-compact" framing and the pre-first-response 80% left fallback are unchanged.
• New sess: N% left until <time> segment from rate_limits.five_hour (the 5-hour session window): used_percentage drives the remaining %, and resets_at (a Unix epoch) is rendered as the window's reset time in local computer time, e.g. sess: 64% left until 2:30pm. The 12-hour clock is formatted by hand (hour % 12, am/pm) to stay portable across the mac and linux brew builds; the until clause is dropped when resets_at is absent. White normally, yellow at <=20% left. The whole segment is absent on API-key billing and before the first response, where rate_limits isn't present.
• Width-aware MCP layout: read COLUMNS (set by Claude Code v2.1.153+) and wrap the MCP server list onto continuation lines instead of letting it truncate at the terminal edge. Stays on one line when it fits; otherwise packs servers after the inline MCP: title and spills overflow to new lines, dropping the whole block to its own line when even the first server won't fit. Width is measured on visible text with ANSI escapes stripped, minus a 2-column margin for the double-width emoji and the right-side notification area.
• Sort MCP servers connected-before-failing, each group alphabetical, so the list is stable and the broken ones cluster at the end.