Add experimental Jev semantic observer boundary

[halqme]

Summary

• add packages/semantic-predicate as a Pi-independent workspace package
• call Jev through OpenRouter's native Decisions API (POST /api/alpha/decisions)
• model the native Jev primitives and response shapes: Noul, Choice, and Score
• add extensions/semantic-observer as a thin, explicitly invoked Pi adapter
• derive Jev state from Pi Kit runtime evidence instead of caller-authored summaries
• keep all semantic observations advisory; they cannot mutate task, satisfy verify, or unlock completion

Boundary

task / context / code / edit / verify runtime evidence
                    |
                    v
extensions/semantic-observer
  - selects a predicate-specific evidence projection
  - asks focused semantic questions
                    |
                    v
packages/semantic-predicate
  - no Pi dependency
  - OpenRouter Decisions API
  - typed Jev questions / answers

extensions/* and packages/* are both root workspaces. The evaluator package knows nothing about Pi.

Pi runtime evidence

The caller now supplies only the observation IDs to run:

semantic_observe({
  observations: ["scopeDrift", "verificationGap", "consistencyRisk"]
})

It does not supply request, changes, verification prose, or repository context.

A new side-effect-free task/evidence.ts packet exposes the same runtime facts used by task.review_context without creating a review request:

• task goal and acceptance criteria
• latest checkpoint / current stage
• tracked read/context/code/edit/write resource provenance
• workspace delta since the task baseline
• verification evidence and provenance

The semantic observer augments that packet with:

• a bounded Git diff from the task's captured baseline
• bounded excerpts from the exact successful read / context tool results already seen in the session, joined back to resource events via toolCallId

The working plan is included only as current-stage context and explicitly marked as a hypothesis, not as authority to expand task scope.

Context projections

scopeDrift
  task goal + acceptance
  + current checkpoint
  + tracked mutations + task-baseline diff

verificationGap
  task goal + acceptance
  + tracked mutations + task-baseline diff
  + executed verification only

consistencyRisk
  tracked mutations + task-baseline diff
  + observed repository paths
  + bounded excerpts from the exact read/context results already observed

The context policy remains:

1. Filter before inference.
2. Prefer runtime facts and repository evidence over model-authored summaries.
3. Give each judgment only the evidence it needs.
4. Keep deterministic facts in code.
5. Preserve Jev probabilities; thresholds and actions remain outside the observer.
6. Do not automatically accumulate broad session history, repository dumps, or prior Jev outputs.

Jev API

• endpoint: https://openrouter.ai/api/alpha/decisions
• request: model + state + questions
• default model pinned to typesafe/jev-1.13
• Noul returns noul as the yes-probability
• Choice returns choice + probabilities + confidence
• Score returns score + legend + probabilities + confidence

Safety / scope

• explicit call only; no automatic hooks
• no changes to task, verify, repository, or delegate authority
• no action is taken from a probability
• missing OPENROUTER_API_KEY only fails the observer call
• no caller-authored semantic evidence payload
• no broad session or repository context is sent by default

Validation

• semantic-predicate tests cover native response parsing and Decisions API request shape
• semantic-observer tests cover the three runtime evidence projections
• task review-context tests continue to exercise the shared evidence packet path
• root workspace CI runs lint, typecheck, and tests on Linux/macOS