Feature: distinguish human vs. scheduled turns on the wire (turn trigger)

[akandola]

Problem

Consumers of the event stream can't tell whether a turn was human-initiated or
scheduled/event-driven. Both paths converge on the same runtime route — the scheduled
path (ProxyChannel.fireTurn) POSTs to /conversations/:id/messages, the same route a human
message uses — with no marker distinguishing them. acting-user doesn't cleanly separate
them (single-player humans have none; legacy routines lack acting-user). So there's
currently no way, on the wire, to know a turn's origin.

Why it matters

For observability/analytics consumers, distinguishing autonomous (scheduled/event-driven)
runs from human-initiated ones matters — e.g. reasoning about how much an agent does
unattended. Today that's invisible downstream.

Proposed approach (open to your preference)

Introduce an internal turn-source signal and surface it on the wire. One option: a
x-houston-turn-source: human | scheduled header set by each channel profile (the
scheduled path → scheduled; the human path → human), threaded through start-turn /
TurnRequest.source, and emitted — e.g. as a third field on the turn_start event we're
adding in the PR below, or its own frame.

This touches ~6 files across host + runtime and introduces a new internal convention, so
per CONTRIBUTING we're raising it as an issue first before implementing. Happy to build it
whichever way fits best — a header, a request-body field, a TurnRequest.source, etc. What
mechanism aligns with how you think about the turn path?

Pairs with the PR we're opening now that adds turn_start with provider + model; trigger
is the natural third field but needs this cross-layer plumbing, so we split it out.