feat(widget): add Cache Hit / Read / Write widgets

[Viperwow]

Summary

Adds a Cache widget category that exposes prompt-cache efficiency. Today Tokens Cached sums cache_read_input_tokens (served from cache, the cache hit) and cache_creation_input_tokens (written to cache this turn) into one opaque number, so there's no way to see your cache hit rate or how much context is being re-written cold. These widgets surface exactly that, to help users measure and maximize cache reuse.

New widgets

Each widget toggles between per-turn ("last action", default) and session scope via the t keybind (mirrors the existing context-inverse used/remaining toggle and the speed-widget config pattern).

Widget	Example	Meaning
Cache Hit	Cache Hit: 88.0%	read / (read + creation) — share of cacheable context served hot vs rewritten cold
Cache Read	Cache Read: 88.0k (84.5%)	cache-read tokens + share of context
Cache Write	Cache Write: 12.0k (11.5%)	cache-creation tokens + share of context

Inline: Cache Hit: 88.0% │ Cache Read: 88.0k (84.5%) │ Cache Write: 12.0k (11.5%)

Approach

• No new data source. Per-turn values come straight from the live status JSON (context_window.current_usage); session values are summed from the transcript in getTokenMetrics.
• TokenMetrics gains optional cacheReadTokens / cacheCreationTokens (a split of the existing cachedTokens, which is unchanged — the Tokens Cached widget and its tests are untouched, fully backward compatible).
• Shared helpers keep it DRY: parseCurrentUsageTokens (reused by getContextWindowMetrics), cache-scope.ts (scope toggle), cache-metrics.ts (formulas + tokens (pct) formatter).
• Defaults: Cache Hit/Read green, Cache Write yellow (recolorable in the TUI). Widgets return null on missing data / zero denominator, like existing widgets.

Testing

• New src/widgets/__tests__/CacheWidgets.test.ts (turn/session scope, raw mode, null/zero-denominator edges, scope-toggle keybind, preview).
• Updated jsonl-metrics tests for the read/creation split.
• bun test (new + updated suites green) and bun run lint (tsc + eslint, zero warnings) both pass.

[Viperwow]

Hi,

@sirmalloc I've been using your project for about a week and am pretty happy! Thank you for your work!

In the meantime, I've been struggling to measure cache hits during experiments with different tool setups, and that's my PoC for widgets to implement to address the issue. Co-authored with AI, but tests seem to be working, covering happy paths, and I've reviewed the code part too manually.

I hope this PR could be a good starting point for an improvement or even be merged.

[Viperwow]

I know that there's another way of implementing this without modifying the source code by adding custom command execution. I haven't used the recommended https://github.com/ryoppippi/ccusage and cannot tell whether it can show only the values for those items to implement it as a custom Widget.