Wrap the status line rather than drop the per-model limit

The measured ladder tried every single-line variant before considering a
second line, so whenever line one plus the full group did not fit — from
around 92 usable columns upwards, depending on how long the branch, cwd and
model names are — it emitted rl_bare (5h and 7d only) and silently dropped
the per-model (Fable) bar that is the main reason the group is worth
rendering. On a 110-column laptop the limit was invisible.

Reorder so a second line beats losing that bar: one line rich/mid/lean, then
wrap, and rl_bare only when WRAP_NARROW is false. Reset times and
extra-usage credits are still given up rather than wrapped for, which makes
the rendered content non-monotone in width; the comment on the ladder spells
that out.

Also stop the tests writing fixtures to the caches the live status line
reads. ~/.claude/statusline.sh is a symlink to the script, so a test run was
visibly rendering fabricated usage — a Fable bar at 10%, extra-usage credits
of $1234.56/$2000.00 — in whatever session happened to be open, and a run
killed before its trap fired would have left that in place for an hour. The
cache directory is now $STATUSLINE_CACHE_DIR (default /tmp/claude) and each
suite points it at a temporary directory, which also removes the
backup/restore dance and the deletion of the live git-status cache.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

claude/tests/run.sh

@@ -4,8 +4,9 @@
 # Usage: claude/tests/run.sh
 #
 # The tests render the status line directly, so they need the same tools it
-# does (bash, jq, git, python3). They briefly replace the cached usage-API
-# response in /tmp/claude with fixtures and restore it on exit.
+# does (bash, jq, git, python3). They point STATUSLINE_CACHE_DIR at a temporary
+# directory and write their fixtures there, so the caches the live status line
+# reads are never touched.
 cd "$(dirname "${BASH_SOURCE[0]}")" || exit 1
 
 status=0