Support fetching prompts from Langfuse prompt management

[AmitAvital1]

Background

Extra already integrates Langfuse for tracing via LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY, no YAML (observability/providers/langfuse/provider.py). Langfuse separately offers prompt management: a prompt has a name, each edit creates an immutable version, and a label (default production) is a movable pointer to one version.

Today prompts.system / user / orchestrator in agents.yml are always a local file path. This issue adds a second source: fetch by name from Langfuse instead.

Proposed YAML — needs research before locking in

# unchanged — bare string still means local file
system: "prompts/orders/system.md"

# new — object form opts into Langfuse (same shape for system/user/orchestrator)
system:
  langfuse:
    name: "returns-system"   # required
    label: production        # default; mutually exclusive with version
    # version: 3              # pins to one immutable revision

Two things the assignee needs to research and settle, not assume: whether production or latest (auto-tracks most recent save, no promotion step) is the better default label for this project; and whether langfuse: as a direct key is right, vs. reserving it for later if a second remote source ever shows up (default to the direct key — don't build for a hypothetical second backend now).

Runtime: fetch once at startup, no caching

One fetch per node when the graph is built at server startup, held for the process lifetime — no TTL, no refetching, no live updates in v1. A fetch failure at startup is fatal: the server refuses to start rather than run with a missing prompt (fail closed).

generate: health check, not scaffolding

#18 makes generate create local stub files for missing prompts — that doesn't apply here, since a Langfuse-sourced prompt has no local file by design. generate should skip these nodes for scaffolding, and instead run a lightweight connectivity check (credentials present, package importable, Langfuse reachable) — not a per-prompt existence check.

Non-goals for v1

• Chat-typed Langfuse prompts (role-tagged message arrays) — text-type only; error clearly on chat-type rather than mis-render it.
• Caching with a TTL, so a promoted prompt change is picked up without a restart.
• Reacting to prompt changes in Langfuse (e.g. refresh on label move) instead of requiring a redeploy.