One trace_id, both lanes

The broker (pramaan-functions src/clerk_worker/runner.py) already injects W3C trace context as meta.traceparent on the agent_invoke.v1 envelope, and the LLM lane already adopts it: agents/atlas/observability/identity.py's agent_invoke_observability_context extracts the carrier and attaches it before contextvars.copy_context(), so the copied brain-thread context carries the remote parent (Story-1439 / 1463 proved the Logfire-lane spans join it). The GAP: nothing taught the GENERAL lane the same trick, and the run_id dimension landed on a general-lane span only when a manual_span happened to pass it explicitly.

1 · The one missing wire — the enricher on the global provider

The LLM lane gets the converged AmbientIdentitySpanProcessor on its dedicated provider (in configure_observability). The general lane — the ADOT-owned global TracerProvider — never got it. observability/identity.py adds the one wiring function that fixes that, without ever owning or overriding the global provider.

# agents/atlas/observability/identity.py
def configure_general_lane_identity(provider=None) -> bool:
    # enrich the EXISTING global provider; never construct, never set_tracer_provider (ADR-104)
    target = provider or _otel_trace.get_tracer_provider()
    add = getattr(target, "add_span_processor", None)
    if not callable(add):           # API default proxy (no ADOT) -> clean no-op (local/CI)
        return False
    add(AmbientIdentitySpanProcessor())   # the SAME single converged enricher the LLM lane uses
    return True                     # idempotent (warm-container guard), never raises

Now run_id / firm_id / form_id / model_id / agent.key land on every general-lane span at on_start — including ADOT auto-instrumented spans the brain never touches — sourced exactly as ADR-124 §D10 requires (run_id/form_id from allowlisted baggage; firm_id from the process-local ContextVar, never baggage). main.py calls it once at cold start, unconditionally (not gated on LOGFIRE_TOKEN): the general lane is correct even with the Logfire lane off.

2 · Why trace continuity already holds in both lanes

It is the same attach for both. runtime.py stream_invoke wraps the context capture so the remote parent rides into the brain thread; the general-lane manual_span (via the global tracer) and the LLM-lane Pydantic-AI span both create children of that one attached parent — so both inherit the upstream trace_id.

# agents/atlas/runtime.py — ONE extraction, both lanes ride it
with agent_invoke_observability_context(meta, model_id=config.model_id):
    ctx = contextvars.copy_context()   # remote parent now in the copied brain-thread context
brain = threading.Thread(target=lambda: ctx.run(_pump), ...)   # both lanes' spans descend from it

3 · The no-parent fallback is preserved, not replaced

4 · Proof — offline, deterministic, no key, no egress

agents/atlas/tests/test_atlas_observability.py gains four tests mirroring the existing #1439/#1463 harness (real OTel SDK + in-memory exporters + the broker→AgentCore thread hop; no network):

# both lanes share ONE trace_id == the injected parent, run_id on EVERY span
root = ATLAS_INVOKE  # general lane (manual);  auto = adot-auto-span  # general lane (raw tracer, NO explicit run_id)
chat = pydantic-ai chat span  # LLM lane
assert root.trace_id == parent and auto.trace_id == parent and chat.trace_id == parent  # ✓ one id, both lanes
assert auto.attributes["run_id"] == "run-1843"   # ✓ stamped by the enricher, not an explicit pass

# no inbound traceparent -> general-lane root keeps the X-Ray-format fallback
_assert_xray_format_trace_id(root.trace_id, captured_at=now)   # ✓ epoch-prefixed, not W3C-random

The adot-auto-span is the sharp proof: it is created through the raw global tracer with no explicit attributes, so its run_id can come only from the enricher on the global provider — exactly the gap this Story closes. Two focused tests pin the wiring itself (added once, idempotent; clean no-op when ADOT is absent / the provider has no add_span_processor).

5 · Verify gate

ruff clean on agents/atlas and tests/ scripts/. pytest 618 passed in agents/atlas (4 new + existing) and 146 passed governance. brain-egress R12 gate clean — no new egress, no baggage write outside the helper, no set_tracer_provider, no provider construction or X-Ray-SDK outside the bootstrap. Observability ONLY: the user-facing answer path is byte-identical (pramaan-docs ADR-110 thin brain).