[omartelo]

Summary

The footer now says how much of the last turn's input came back from the prompt cache:

• Cache: 98% hit, green from 50% and amber below, so a turn after the cache expired (0%, amber) stands out as the one billed at the full rate.
• A bar under it splits the turn's input into cached (muted), written to the cache (amber) and uncached (foreground).

It sits in the context window tooltip on Claude Code and Codex, and in the cost tooltip on oh-my-pi and opencode, which record the split but no window.

Also fixes the Codex cost billing a cache write twice: cache_write_input_tokens is part of input_tokens (Codex fills it from the Responses API's input_tokens_details; its own parses_cache_write_token_usage test has 40 cached + 60 written of 100 input), but it was priced at the cache-write rate and again as uncached input. Cost and the cache bar now share one split, codexCache.

How

• api.SessionUsage gains Cache *SessionCache {read, write, uncached}; contract.golden moved with it.
• Claude Code: the same assistant line the ring reads (cache_read_input_tokens, cache_creation_input_tokens, input_tokens). A compaction drops the split until the next assistant line, since the turn before it no longer describes the window.
• Codex: last_token_usage, both cache counters taken out of input_tokens.
• oh-my-pi: the newest assistant line's usage.input, cacheRead, cacheWrite (omp already takes the cache out of input, calculateOpenAIUsageAccounting).
• opencode: the newest finished assistant message row of the session, tokens.input, tokens.cache.read, tokens.cache.write; a sub-agent's own session is not read.
• Native: footer_cache.go draws the line and the three-part bar, under the context token line or under the cost lines when there is no window; one new catalog key, shell.footerSession.cacheHit, in all seven locales.

Provider gap (Invariant 5)

Crush folds its cache counts into its totals, Kiro files every token count as zero, and Cursor and Antigravity have no usage reader, so the four show no cache line. Written down in docs/ceilings.md, the Provider support table of every README, and docs/providers/.

Tests changed

• TestCodexTranscriptCostPricesTheRunningTotal asserted the double count, which Codex never promised: its uncached share is now 575 of 1000 input, not 600.
• TestSessionUsageReportsTheTranscriptTail, TestSessionUsageReportsCodexRollout, TestScanCodexContextUsage: the event contract gained Cache.

Test plan

• go test ./internal/terminal ./api: split per provider, compaction clears it, opencode skips an unfinished turn and a sub-agent's, a missing message table reads as no line, the cost-only event carries it, Codex cost with a cache write.
• go test ./cmd/lich-native: decode keeps a valid split and drops a malformed or negative one; cacheHit rounding and the empty cases.
• Local gate (backend and native/): gofmt, vet, test.
• Rendered both tooltips headless, dark and light: context at warm (98%), mixed (62%) and cold (0%), cost on a provider with no window.
• The opencode query run read-only against a real opencode.db.