feat: price the token volume at API list rates

[AqueGen]

Tokens need translating; dollars do not. tokens now prices the volume it already counts.

At API list prices (rates as of 2026-08-11), this is what the subagent volume above would have cost on the Claude API:
  as it ran                                            $2,365
  had every subagent inherited its session model       $4,402
  difference                                           $2,037
  main sessions, same rates (not routable)            $15,560

Why this is not the dollar fiction the file warns about

The old note said "no dollar fiction", and the objection it aimed at was never the translation - it was the saving. A saving needs a baseline nobody measured, which is why the README already calls the volume chart an upper bound rather than a measurement. So there are two figures and a difference, all three labelled, and the difference inherits that same caveat verbatim rather than being presented as money recovered.

It is also a counterfactual by construction: on a subscription none of this is paid. The wording says so on its own line.

The correctness story is per token type

Base input, cache reads at 0.1x, five-minute cache writes at 1.25x, one-hour writes at 2x, output at its own rate. These are an order of magnitude apart and cache reads dominate real transcripts, so a flat volume multiply would overstate the bill several times over. Transcripts carry cache_creation.ephemeral_5m_input_tokens and ephemeral_1h_input_tokens, so the write split is exact; a line with only the flat total takes the cheaper five-minute rate rather than guessing upward.

Prices are enumerated per model rather than per family, for the reason EFFORT_SUPPORT already is: families span price changes, Opus 4.1 bills at three times Opus 4.5, and a loose pattern would quietly misprice a retired model. Anything absent from the table is excluded from every figure and declared on its own line - never counted as free. Sonnet 5 is the one model whose rate changes on a calendar date rather than with a new id, so a window is priced at the rates in effect at its end, not at today.

What biases the number, and which way

Downward, both of them, which is the direction to err in:

• models from Opus 4.7, Sonnet 5 and Fable 5 onward use a tokenizer producing about 30% more tokens for the same text than Sonnet 4.6 and earlier, so re-pricing a cheap model''s token count at an expensive model''s rate understates what that work would really have cost there;
• cache writes on lines with no TTL breakdown take the cheaper rate.

Tests

64 pass. Five are new: every token type at its documented multiplier asserted to the cent ($9.35 on haiku, $46.75 for the same counts on opus-5), the untyped-write fallback, unpriced volume excluded and declared, the section absent when nothing in the window can be priced, and the price stamp reaching the output instead of sitting in a comment where nobody checks it.