BYOC protocol cleanups (from SDK implementation)

Six rough edges in BYOC's wire surface, surfaced while implementing a
Python pipeline SDK against the existing protocol. None blocking;
flagging for backlog.

Protocol cleanups

1. Inconsistent route prefixes. Same operation appears as /process/stream/* (gateway), /ai/stream/* (orch), /stream/* (runner). Unify or rename — the /ai/ prefix is historical baggage from when AI workers were a separate path.

2. Trickle control channel bypass for /update is a permanent workaround. byoc/stream_gateway.go has a commented-out controlPub := stream.ControlPub with a "broke on base64 binary" comment. Either fix trickle's binary handling or formalize HTTP as the control plane and remove the trickle control channel from the protocol.

3. /ai/stream/payment leaks payment concern into the protocol surface. Separate endpoint just for payment forwarding feels misplaced. Bundle into existing requests (header-based, like Livepeer-Balance) or document as orch-internal so SDK authors don't think they need to model it.

4. Capability registration as an out-of-band dance. Every BYOC compose ships a register_capability sidecar that POSTs to the orchestrator with a separate auth scheme (orchSecret Bearer). Self-registration from the runner with heartbeats would be cleaner — no sidecar, no orchestrator-must-be-up-at-runner-startup constraint.

5. No explicit backpressure semantics in trickle. Publisher writes, subscriber reads; a slow subscriber just causes queue/drop with no protocol signal back to the publisher ("slow down" / "buffer full"). Critical for real-time A/V quality at scale.

6. Overlapping BYOC route prefixes. /process/request/* (batch BYOC), /process/stream/* (live BYOC), /ai/stream/* (orch internal) — three operations sharing the BYOC namespace via two different prefixes. Worth considering more orthogonal naming for a v2 protocol. (Old non-BYOC batch path is deprecated and ignored.)

Implementation polish (small fixes)

Smaller code-level cleanups uncovered while implementing the SDK against the existing wire protocol. Filed as separate issues so they can be picked up independently:

• #3922 — BYOC gateway: ffmpeg output process logs spurious "Broken pipe" / "Conversion failed!" on every /stream/stop. Same pattern duplicated in server/ai_live_video.go. See issue for two proposed approaches (low-risk noise suppression vs. shutdown-sequencing fix).

Context: implementation experience from livepeer/livepeer-python-gateway#8. Happy to break out any of the protocol items above into individual issues if useful.

Begin by selecting one of the six protocol items; the issue names byoc/stream_gateway.go for the /update workaround and server/ai_live_video.go for related shutdown logging. No test files or single entry point are identified, so done requires a separately scoped change with explicit protocol behavior and validation.