Web dashboard Cache read/write cards sum entire history instead of current period

[Llldmiao]

Bug Description

The web dashboard (codeburn web -p 30days) shows incorrect values for the Cache read and Cache write stat cards. They display totals from the entire history.daily array (up to 365 days of backfill), rather than only the data within the user-selected period (e.g. 30 days).

Steps to Reproduce

1. Run codeburn report -p 30days in the terminal
2. Run codeburn web -p 30days and open the web dashboard
3. Compare the "cached" value in the terminal with the "Cache read" card in the web UI

Expected Behavior

Both views should show the same Cache read value for the same period (30 days).

Actual Behavior

• Terminal (report -p 30days): shows 1391.1M cached
• Web (web -p 30days): shows Cache read: 2.07B

The web value is ~1.5x larger because it sums over ~365 days of history data.

Root Cause

In the web frontend (dist/dash/assets/index-*.js), the Cache read/write cards compute their values by reducing over the entire e.history.daily array:

l = (e?.history.daily) ?? []
o = l.reduce((b,x) => b + x.cacheWriteTokens, 0)  // Cache write card
c = l.reduce((b,x) => b + x.cacheReadTokens, 0)    // Cache read card

However, history.daily is populated from BACKFILL_DAYS = 365 days of cached data (for the trend chart), not limited to the current period. Meanwhile, other cards (Cost, Calls, Sessions, Tokens in/out) correctly use e.current which is properly scoped to the selected period.

Suggested Fix

The frontend should use e.current.tokens.cacheRead and e.current.tokens.cacheWrite (which are correctly scoped to the period in buildMenubarPayload) instead of summing history.daily.

Alternatively, filter history.daily to only include dates within the selected period before summing.

Environment

• codeburn version: 0.9.14
• OS: macOS (Darwin 25.3.0)
• Node: v24.13.0

[ozymandiashh]

Thanks for the precise write-up, the root cause was spot on and made this quick to pin down.

Confirmed: the Cache read / Cache write cards reduced over the full history.daily backfill (BACKFILL_DAYS = 365), while every other card reads the period-scoped current block. The shorter the period the worse the over-count, on my data today was inflated ~197x, 7 days ~5.2x, 30 days ~1.2x.

One wrinkle on the suggested fix: current didn't actually expose the cache token counts yet (it only carried cacheHitPercent, derived from them), and its shape is flat (current.inputTokens, …) rather than current.tokens.cacheRead. So the fix threads cacheReadTokens / cacheWriteTokens from the period-scoped PeriodData through buildMenubarPayload into current, then points both the single-device and the combined multi-device cards at them. The trend chart keeps using history.daily. After the change web -p 30days matches report -p 30days exactly.

While in there I found the same bug on the CLI side (codeburn devices / summarizeOneDevice), which showed period-scoped cost/tokens but all-time cache, so that path is fixed in the same PR.

Fix in #586.