feat(gateway): streaming proxy with budget enforcement (OpenAI)

[prashar32]

The most-requested adoption gap: stream:true on POST /v1/chat/completions. Closes #22.

Behavior

• Budget enforced before the stream opens — BeginStep + CanProceed; if the run is over budget, a 402 is returned and the provider is never hit.
• Authentic SSE pass-through — the provider's SSE chunks are forwarded to the client verbatim (no parse-and-reserialize), so an OpenAI client sees a real OpenAI stream. Token usage is teed off the final usage chunk for metering.
• Live-stream cut — the call runs under the run's context, so the time budget, kill switch, or client disconnect interrupts a stream in flight.
• Metered at stream end — the streamed call is recorded to the cost ledger, so the next call is refused if this one went over budget.

Honest scope: OpenAI reports usage only in the final stream chunk, so true mid-stream dollar/token cutting isn't possible — those are pre-checked + post-recorded; the time budget / kill switch are what cut a live stream. A new optional provider.Streamer interface gates this; OpenAI implements it. Providers without it, and the Anthropic /v1/messages endpoint, return a clear 501 rather than silently buffering — native Anthropic SSE streaming is the documented next step.

Tests

• Provider: mock SSE upstream → chunks forwarded verbatim, usage/model parsed (incl. the stream+include_usage request shape).
• Gateway: client receives the SSE and the call is metered to the run; a second streamed call is refused at 402 before the provider is hit (loop budget); a non-streaming provider returns 501.
• gofmt/go vet clean; full suite green; race detector clean on the gateway + provider streaming path.

---

Built alongside (parallel, in isolated worktrees) #122 (shell completions) and #123 (/metrics); all three touch disjoint files. CHANGELOG entries will need trivial merge ordering.