feat: interaction log for run/node prompts + outputs (follow-up to #722)

[rafeekpro]

Summary

Extend the EU AI Act interaction log (#722, PR #784) to the bigger compliance surface: agent/node prompts and outputs from pipeline runs.

The mechanism already exists and should be reused as-is:

• dap_engine.redaction.redact() — exact-value scrub (decrypted instance env vars) + pattern scrub
• persistence/interaction_log.record_interaction(surface="run"|"node", ...) — the table, admin API (GET /interactions?surface=...), retention purge and config all already handle non-assistant surfaces

Scope

• Record a redacted interaction per node execution that calls a model: the rendered prompt (redacted_request) and the raw model output (redacted_response), with provider/model/tokens_used from the adapter result
• Decide the integration point: execution/node_executor.py (has the prompt + output + adapter metadata) vs. the adapter base (closer to the wire, catches CLI runtimes too)
• known_secrets resolution inside the run context — runs resolve instance env vars already; thread the decrypted mapping to the redaction call
• Volume control: node outputs can be large (node_output_chunks exists for streaming) — decide truncation policy for the interaction record (e.g. cap at N KB with an explicit [TRUNCATED] marker in extra)
• Tests: redaction of secrets embedded in prompts and in model outputs; volume cap; per-surface filtering in the admin API

Non-goals

• Dashboard admin view (tracked separately if wanted)
• Changing the assistant path (done in feat: EU AI Act interaction log — redacted assistant transcripts, admin API, retention (#722) #784)

Mechanism from #722; this issue is the "Follow-up: agent/node prompts + outputs" item from its acceptance criteria.