fix(worker): attribute learning to project/team/role and commit the prompt snapshot

[GQAdonis]

Summary

Plan PR A2 of the team-aware learning work (design §6, §8 U2).

• Attribution. LearningJob gains optional projectId, teamId and roleId, all serde(default), so jobs queued by older hooks still parse. enqueue_memory now derives the scope keys from them:

  • user_id is the project id, falling back to project_scope();
  • agent_id is <team>/<role> when both are set, otherwise @project;
  • shared-scope learning uses @global for both.

  Previously agent_id was always null, which is the source of the unattributed records.

• Legacy task steps. The legacy add_task_step form ({stream, description}) now gets non-null agent_id and user_id keys (PROMETHEUS_PROJECT_ID, or an explicit project:unknown that the surreal-memory re-key can repair).

• Hashed payloads are untouched on purpose. Already-normalized payloads (add_memory, stream_name task steps) are bound by their stored payloadHash and any server receipt, so rewriting them would break reconciliation. New operations get correct keys at the source, the skill-pack memory bridge.

• Prompt snapshot. process_job now commits it after the upsert. Without that, session records were written to disk but were invisible to pk context.

Test plan

• cargo test -p pk-learning-worker --test attribution, 2 tests:
  • attributed and unattributed jobs produce tlm-fixture/api-dev and @project keys, no queued operation has a null key, and both session entries are in the committed project snapshot;
  • a legacy task step is submitted with agent_id=@project and user_id=$PROMETHEUS_PROJECT_ID, captured at a fake ledger.
• cargo test -p pk-learning-worker --test worker: 5/5 still pass.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.