[usage] Account for what a call charged and what it cost

[yusiwen]

Closes #181. Roadmap entry A3's cost half (tracker: #168); the price-table question stays open in #182.

What changed

• S1 — the lanes. types.Usage gains CachedPromptTokens, CacheWriteTokens, ReasoningTokens:
  one field per billing lane, so OpenAI's nested prompt_tokens_details and DeepSeek's flat
  prompt_cache_hit_tokens land in the same place. The nested shape wins when a route sends both, so
  nothing is counted twice.
• S2 — the charge a route reports. types.Cost, CostSource, CostEvent, CostTotals and
  ChatResponse.Cost; the OpenAI-compatible provider reads cost, falling back to
  cost_details.upstream_inference_cost for a BYOK request that billed nothing, and carries the
  amount verbatim. SetRouteInfo names the route and the unit it bills in.
• S3 — the rates a user declares. pricing in the configuration, keyed by <route>/<model> with
  * in either position, resolved most-specific-first; PriceTable.Amount carves the detail lanes out
  of the totals and falls each detail rate back to the lane it belongs to.
• S4 — display and docs. The status bar shows cost: <totals> plus (+N unpriced); CHANGELOG.md,
  CODEBASE.md, docs/roadmap.md, config.example.json.

Why

A3 left "what did this cost" unanswered for two independent reasons, and they need different
mechanisms:

• A route may report the charge. OpenRouter returns cost (and a prompt_tokens_details split of
  cached and cache-written input) in its usage object, so no price list is involved at all. That
  number is a reading of the bill, so it is carried verbatim — never recomputed, never added to a
  figure derived from a table.
• Otherwise a rate has to be applied, and a rate needs lanes. DeepSeek — the default provider —
  bills four input lanes (cache hit vs cache miss) times a peak/off-peak rule: a 50× spread on
  deepseek-flash input alone. A single input price cannot express it, which is why the usage detail
  comes first and why a naive two-number table was rejected in [usage] Decide whether to vendor a model price table (and whose data it would be) #182.

Two deliberate refusals, both pinned by tests:

• A call whose tokens we estimated is not priced. Multiplying a declared rate by a guessed token
  count would dress a second estimate as a measurement.
• An unpriced call is counted as unknown, never as zero. Ollama and subscription plans have no
  per-token price at all, and 0 would read as "free" in exactly the budget decisions this feeds.
  The bar shows the total and the unknowns beside it, because a total without them is a different
  claim.

Verification

• go test ./... -count=1 → 13 packages ok, 0 failed. go test ./... -count=1 -race → same.

• make staticcheck (v0.8.1) → exit 0; go vet ./... and gofmt -l . clean.

• The 27 committed frame goldens pass unchanged: the cost segment appears only once a cost is known.

• Eight mutations of the new behaviour, each failing on an assertion:

  Mutation	What the test said
  nested usage detail ignored	CachedPromptTokens:0 … want CachedPromptTokens:60 CacheWriteTokens:10
  upstream cost ignored	cost = <nil>, want 19 USD
  lanes added on top of the totals	amount = 16.1, want 11.1 (and 19, want 14 for the fallback)
  a wildcard outranking an exact key	Lookup("openrouter/claude-sonnet-4").InputPerMillion = 3, want 4
  an estimated usage priced anyway	events = [{… Source:declared}], want one unknown
  an unknown cost recorded as zero	CostTotal = "0", want empty
  a declared rate outranking the route's figure	CostTotal = "200 USD", want "7 credits"
  the UI dropping the unpriced count	bar = "… cost: 0 …", want an explicit unknown

• Money is computable by one command, which is this repository's rule for a number in a document:
  TestPriceAmountCarvesTheLanes fixes a usage and a rate and asserts the amount, lane by lane.

Honest scope

• No live provider call was made. The reported-cost fixtures mirror OpenRouter's documented usage
  object; the lane fixtures mirror OpenAI's and DeepSeek's documented fields.
• A route that reports a charge but no currency code is displayed without a unit rather than labelled
  with a guessed one (cost_currency names it; it is not defaulted).
• Peak/off-peak pricing is not modelled: a declared rate is one rate. A provider that prices by the
  hour needs either a reported charge or a rate the user updates, and saying so is better than
  averaging the two.
• Pricing a sub-agent's call is done with the same table, but its spend still is not attributed to the
  parent's totals — that is roadmap G1, as in [usage] Stop a run at a cumulative token budget (per run and per session) #179.