v2.4.4 — token-throughput display + metric-correctness fixes

[soumyadebroy3]

Summary

The menubar/HTML/CLI reported token usage as input + output only, while cost billed all four token buckets. On cache-heavy agentic runs that made the headline "tok" ~2–5% of real throughput (a ~1.1M-token workflow read as ~57K) and cost-per-"token" look absurd. This release puts the token figure on the same basis cost already uses — total throughput = input + output + cache read + cache write — and fixes a sweep of related metric-correctness issues. Bumps version to 2.4.4.

Token display (core fix)

• Hero, trend, and HTML headline now show total throughput, matching the Anthropic usage object and the Claude Code / Warp counters. Cache volume is forwarded in the menubar payload (current.cacheReadTokens / cacheWriteTokens).
• Applied across macOS (Swift), Windows (React), and GNOME (JS).
• Hero ↑/↓ split now follows the conventional in↑ / out↓ direction; effectiveTokens proxy weights cache-write at 1.25× (was 1×).

Metric-correctness

• Cache-hit rate uses one denominator everywhere (input + cacheRead + cacheWrite); menubar/HTML previously dropped cache writes and disagreed with the TUI.
• Trend "% vs prior" tracks the token delta when tokens are shown (was always cost-based).
• "1-shot" rate exposes its sample size (e.g. 0/1) instead of a bare "100%" over a hidden denominator of 1.
• "Yesterday" shows "—" for a no-data day instead of a false "0".
• Last 7 Days / Last 30 Days span 7 / 30 calendar days, not 8 / 31.

Forecast / leverage / retry

• Forecast "On pace for" projects from completed days only (today's partial day no longer drags it low); "Avg/day" + week-over-week use completed-day windows; day 1 no longer false-fires the overspend tip.
• Leverage is period-normalized (single-period value scaled to a 30-day run-rate before comparing to the monthly price); zero-paid 999× sentinel removed; HTML banner names the actual window.
• Retry detection counts a re-edit after any intervening check (not only Bash) while not flagging sequential multi-file edits; retry-tax labeled a worst-case estimate.

Pricing

• claude-opus-4-8 added to MANUAL_ENTRIES ($5/$25 per M) so offline builds don't fall back to the legacy claude-opus-4 rate ($15/$75) and overstate Opus spend ~2.6×.

Not included

• Long-context (>200K) pricing tier — no published Opus-1M premium rate to apply without inventing numbers; intentionally deferred.

Verification

• 788 tests pass (incl. 6 new retry-detection tests; updated cli-date + snapshot-contract tests).
• TS typecheck clean · macOS Swift build ✅ · Windows tsc ✅ · GNOME syntax ✅.
• CLI smoke: today's hero went from a misleading 336K tok to a truthful 30M tok, reconciling with the billed cost (99.9% cache hit).