fix(pricing): a missing cache rate prices as a floor, not a silent zero

[as-koshiyama]

Closes #16.

What

For a model whose LiteLLM entry omits cache_read_input_token_cost or cache_creation_input_token_cost, cache tokens were multiplied by an unwrap_or(0.0) rate and returned as Cost::Known — "unpriced is not free" broken silently inside the headline number, on the token kind that dominates a coding-agent corpus. The CHANGELOG claimed input-rate fallback; code and changelog now agree, per the issue's checklist.

Semantics chosen

The two fallbacks lean opposite ways for the same reason — never overstate:

• Cache creation → input rate. Providers that publish a write rate set it above input, so this understates.
• Cache reads → zero. Their real rate sits below input, so an input fallback would overstate.

Both understate, so the result is a new Cost::Floor(f64): rendered ≥$x in the Cost view, {"state":"floor","usd":x} in --json, and counted into every total's ≥ marker via Cost::not_fully_priced() (which replaces is_unpriced() at each aggregation site — a floor's dollars are real, unlike unpriced's, so the two stay distinct states). A rate listed as 0.0 still prices as zero exactly, because the table said so — only an absent key floors, and only when the usage actually has tokens of that kind.

Also per the issue's last checkbox: the parse-time zero-cost filter now keeps cache-only entries instead of dropping them with the embeddings, and the Cost view's floor title reads "N model(s) not fully priced" rather than "have no price", because a floored model has one.

Verification

The issue's acceptance test verbatim (missing_cache_rates_price_as_a_floor_never_a_silent_zero): cache reads + creations against an entry with input/output rates only → Cost::Floor(2.0), non-zero, not exact, not unpriced. Plus: a missing rate with no cache tokens stays Known; a cache-only entry survives parsing and prices exactly; absent keys parse to None while listed values parse to Some. cargo fmt, clippy -D warnings, all 246 tests pass.

Note for review ordering: independent of every open PR; will need the usual one-line CHANGELOG rebase after whichever lands first.

🤖 Generated with Claude Code

[kleyt0n]

lgtm