fix(quota): use resets_at, not updated_at, as the freshness signal

[arzafran]

What this does

PR #150 session-scoped ~/.claude/tmp/rate-limits.json and took the max
used_percentage across entries fresher than CACHE_STALE_MS (10 min). That
fixed cross-workspace clobber, but the freshness signal it used —
updated_at — is wrong: it records when a statusline render wrote an
entry, not when it observed the reading. A session idle for days still
re-renders its statusline on every keystroke elsewhere and rewrites its
ancient reading with a brand-new updated_at, so it always looks fresh.

Live evidence from a real machine, 8 sessions all "written" within 30 seconds:

session   5h    resets_at        7d    resets_at
6271d111  28%   08-06 13:10 LIVE  75%   08-09 07:00
33276606  50%   08-06 13:10 LIVE  79%   08-09 07:00
9cbb69fa  50%   08-06 13:10 LIVE  79%   08-09 07:00
5005e845  25%   08-06 13:10 LIVE  74%   08-09 07:00
fe277e28   6%   08-03 23:20 DEAD  28%   08-09 07:00
0964fc10  29%   08-04 15:00 DEAD  37%   08-09 07:00
8367e119  67%   08-03 18:20 DEAD  25%   08-09 07:00
071862a0  22%   08-04 15:00 DEAD  35%   08-09 07:00

The derived flat five_hour was landing at 67%, taken from a window
that had already expired on Aug 3 — the correct current-window answer is
50%.

The fix

resolveRateLimits (src/lib/quota.ts) now uses resets_at (already in the
payload, already persisted) as the real freshness signal, per window
independently:

1. drop entries whose resets_at is missing or already <= now (expired window)
2. among survivors, keep only the ones whose resets_at equals the max surviving resets_at (excludes a previous-window reading taken just before rollover)
3. among those same-window entries, take the max used_percentage (usage only rises within one window)

updated_at/CACHE_STALE_MS stays as a secondary guard (bounds map growth,
drops vanished sessions) but is no longer the primary freshness test.
resets_at is parsed defensively — persisted entries are string
epoch-seconds, the parser also tolerates number epoch-millis.

The visible statusline chip is unaffected — it reads the live per-render
rate_limits payload straight from Claude Code, not this cache; only the
cached/derived value consumed by quota-steer.ts, escalate-model.ts, and
Programa's sidebar (ClaudeQuotaSnapshotParser, unchanged flat schema) was wrong.

Test plan

• bun test — 1391 pass, 0 fail (1373 baseline + 18 new cases in tests/quota.test.ts, covering: the exact production regression above, mixed-window exclusion, same-window max-wins, all-expired → null, and string-seconds/number-millis parsing)
• tests/escalate-model.test.ts fixture updated to include resets_at (previously omitted) — the new algorithm needs it to recognize a fixture window as live at all
• bun run typecheck clean
• bun run lint clean
• tests/output-style-preserve.test.ts verified passing in isolation against this commit (3/3) — unrelated failures reported against this suite come from other uncommitted work in the tree, not this change