delivery_observation_payload computes evidence_refs_payload(observation.evidence_events) twice per observation (redundant reverse()/dict work)

[erskingardner]

Summary

For every delivery observation serialized, evidence_refs_payload(observation.evidence_events) — which builds a reverse("api-event-evidence", ...) URL and a dict per evidence event — is materialized twice: once directly for the evidence_refs key, and once inside delivery_observation_state_payloads.

Location

forensics/views.py:2663-2685

def delivery_observation_payload(observation):
    return {
        ...
        "states": delivery_observation_state_payloads(observation),   # builds refs (2678)
        ...
        "evidence_refs": evidence_refs_payload(observation.evidence_events),  # builds refs again (2672)
    }

def delivery_observation_state_payloads(observation):
    refs_by_event_id = {
        ref["event_id"]: ref for ref in evidence_refs_payload(observation.evidence_events)  # 2678
    }
    ...

Impact

On delivery endpoints and the streaming group export, observations with several evidence events pay double the reverse()/dict-construction cost per observation. Prefetch avoids extra SQL, so this is pure redundant CPU (which holds the GIL during export serialization — see docs/deployment.md), not an N+1.

Fix

Compute the ref list once and share it:

def delivery_observation_payload(observation):
    evidence_refs = evidence_refs_payload(observation.evidence_events)
    return {
        ...
        "states": delivery_observation_state_payloads(observation, evidence_refs),
        "evidence_refs": evidence_refs,
    }

Relationship to existing issues

The other "computed twice" issues (#165 message traces, #187 sort, #173 engine_source_values, #257 divergent rollup) are all different call sites; this double-build in the observation payload isn't among them.

Filed by an automated code-review pass.