OIC MVP: end-to-end test of the first investigation scenario #235

Context

The first milestone is one complete workflow: user request → durable investigation → agent exploration → clarifying question → evidence-backed proposals → validated knowledge → queryable API.

Scenario

On a sample repo: "Investigate how this repository's pipeline handles missing input data, identify the relevant business or engineering rules, and determine whether the behavior is intentional."

Acceptance criteria (all must pass)

• Investigation created from the UI and persisted before the agent starts.
• One provider runs it, using tools configured in its runtime.
• Progress and messages are visible in the UI.
• The agent asks a question → waiting_for_user → the user answers in the UI → it resumes, including in a new session.
• State survives a worker restart.
• Structured entity, claim, evidence, and relationship proposals. Invalid output is rejected with useful errors.
• Proposals never write directly to canonical knowledge.
• Accepted claims keep evidence, provenance, scope, and temporal metadata.
• A claim can be revised without losing history. Conflicting evidence can be represented.
• The UI shows a claim with its supporting and contradicting evidence.
• The API returns the claims with no agent session.
• Runs on the remote host with no local machine involved.
• No secrets in logs or checkpoints.
• A retried job makes no duplicate canonical writes.

Tracks: #219 #220 #221 #222 #223 #224 #225 #226 #227 #228 #229 #230 #231 #232 #233 #234