example: demonstrate project-aware coding decisions with Jev and PowerContext

Tracking issue: #1649

Feature description

Provide a reproducible coding-agent example in which PowerContext supplies project history to Jev, Jev makes bounded advisory decisions, and the agent performs and records the actual work.

Priority: P1 — an end-to-end example that users can run on their own project.

Problem and proposed solution

Use a concrete database-compatibility task. A previous session established that SQLite passes but MySQL validation remains outstanding, and recorded a relevant failure or project constraint. In a fresh session, the user asks the agent to continue the fix.

The example should:

1. Persist the original constraint, exact verification evidence, and current Handoff through existing PowerContext operations.
2. Start a genuinely fresh session/process and retrieve bounded project context through prepare_context, resolving only the additional exact evidence needed for assessment.
3. Give Jev the current task plus authorized historical context and explicit questions: does a past failure apply, which supplied requirement still lacks evidence, and is the completion claim supported?
4. Let ordinary host logic and the coding agent choose the next verification action. Jev recommendations do not grant tool permissions.
5. Run the actual checks and record a Task Outcome with exact evidence. Any reusable Experience/Skill proposal follows the existing Review lifecycle.
6. Resume once more to demonstrate that the updated outcome, not the earlier incomplete state, informs the next decision.

Acceptance criteria

• Ship a self-contained example, setup instructions, bilingual explanation, and machine-readable evaluation output. Document live credentials separately from fixture/mock mode.
• Exercise a real PowerContext server and persistent storage across fresh processes. Show the actual PreparedContext bytes/citations delivered and the bounded state sent to Jev, using non-sensitive fixtures.
• Compare current-task-only Jev with the same Jev model supplied by PowerContext history. Add an end-to-end agent baseline without Jev to isolate whether the extra decision step helps.
• Keep task, tools, model versions, and applicable budgets fixed. Measure missed constraints, repeated investigation, false completion claims, successful verification, recovery time, latency, and total token/call cost.
• Include changed requirements, an outdated historical conclusion, no useful history, and a Jev timeout. History and model suggestions cannot override current instructions or authorization.
• Report exactly which decisions, state bytes, checks, and revisions produced the result. Do not imply that prepare_context itself persists a delivery receipt or proves host injection.
• Separate seeded-fixture demonstrations from live task outcomes and publish a recommendation based on observed behavior.

Alternatives considered

An agent can consume PowerContext directly without Jev. That is an important baseline, since a new decision call is only useful if it improves correctness, latency, or cost. A Jev-only current-state example does not demonstrate the value of durable project context.

Additional context

Reuse existing Work Contract, Handoff, Task Outcome, and preparation surfaces. This example can start independently of a production Jev reranker; reuse the evidence-checking adapter when available. Broader cross-host continuity remains in #1352 / #1358, and any new public delivery-receipt design remains in #1356.

• Current public API contract
• TypeSafe state guidance
• TypeSafe guidance on irrelevant state and model limitations

Prepared with AI assistance. This example and its claimed benefits still require implementation and live validation.

Related Jev work: incorporate the advisory evidence evaluator from #1645 when available. This example can run without the production Memory reranker in #1643.