fix(usage): keep session spend accurate without emptying the modal

[Yy-702]

Summary

Follow-up to #611 (already merged). Local /usage could show cached > input (e.g. 144k vs 88k) because per-call / fragment packets were summed with the turn_completed snapshot. A first filter then dropped real turns and the modal went empty mid-turn.

• Sum only turn_completed.
• Ignore fragments with cache and no modelCalls when cache > input (or cache with no input).
• Still accept a real snapshot that omits modelCalls if cache ≤ input.
• Clamp per-turn cache to input.
• Persist spend in sessionStorage so HMR/reload does not wipe totals.
• While a turn is in flight, show that usage updates when it finishes.

Verified locally: 59,780 input / 57,984 cached / 97% hit.

Type of change

• Bug fix
• New feature
• Documentation
• Refactor / chore

Checklist

• I ran pnpm test (sessionSpend + i18n)
• I ran cargo test in src-tauri (no Rust change)
• User-facing strings go through src/i18n/messages.ts (en + zh + zh-TW)
• No window.confirm / prompt / alert for product dialogs
• Docs / docs/llm-wiki updated if behavior changed
• No secrets included

[RongleCat]

Merged, thanks @Yy-702 — session spend now stays on turn_completed snapshots so the modal no longer inflates cache or goes empty mid-turn.