Provider modules already compute cost_usd per turn and emit it on the
JSON-RPC NDJSON wire via hook_streaming
(amplifier_agent_lib/bundle/hook_streaming.py). The chat-completions
HTTP face was throwing this away — extracting only token counts from
kernel usage events. Now it lifts cost_usd through too, accumulating
across sub-turns (a single user turn can drive multiple LLM calls for
tool-call rounds) and emitting the total in the OpenAI usage envelope
as the non-standard `cost_usd` field. Real $$ from the provider's
own pricing, surfaced on the SSE response.

Implementation
--------------
- `_event_translator.extract_usage`: widened return type from
  `dict[str, int]` to `dict[str, Any]`; reads `event['cost']` (set by
  hook_streaming from kernel `cost_usd`) and stamps it on the result
  as `cost_usd: str(...)`.
- `_wire._build_usage_block / stop_chunk / tool_calls_stop_chunk`: new
  `cost_usd: str | None = None` parameter; surfaced on the usage block
  when set.
- `routes/chat_completions`: accumulates `usage_cost: Decimal | None`
  across all usage events in the turn (preserves precision), serializes
  to str on emission, passes through to the terminal chunk helpers.

cost_usd is a string (Decimal precision) and is omitted entirely when
no provider emitted it (older provider modules, third-party endpoints
without cost telemetry, etc.) — standard OpenAI clients ignore the
non-standard field.

Verified end-to-end against the running server: a one-word reply with
claude-haiku-4-5 produces `cost_usd: '0.0118625'` in the terminal
chunk's usage block.

This is a wire translation — it leverages cost telemetry that already
flows on the NDJSON wire. No new pricing catalog or provider-side
plumbing required.

🤖 Generated with [Amplifier](https://github.com/microsoft/amplifier)

Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>