feat(learning): calibrate slop & recommendations vs merge outcomes

[JSONbored]

Parent roadmap: #525
Parent phase: #529

Background

A static rubric can be gamed and never improves. Calibrating slop and recommendation quality against real outcomes is the durable moat. This builds on the (closed) #278 outcome-learning infra rather than reopening it.

Goal

Log slop assessments + recommendations and calibrate against real merged-vs-closed-vs-stale outcomes.

Current Behavior

recommendation-outcomes.ts and outcome dimensions exist (#278 shipped), but slop is not yet calibrated against outcomes.

Desired Behavior

• Lightweight telemetry first: a slop_assessment_events table correlated with PR outcomes.
• Graduate to periodic calibration of SLOP_WEIGHTS and recommendation ranking from observed outcomes.

Implementation Requirements

• Reuse feat(analytics): add outcome learning loop for recommendation quality #278 outcome data; new migration for the telemetry table. No PII / no private context in telemetry.

Public/Private Output Boundaries

• Telemetry is internal; never surfaced publicly; no wallet/reward/trust fields.

Acceptance Criteria

• Assessments + outcomes are recorded; a calibration job can adjust weights deterministically and reproducibly.

Testing Requirements

• npm run test:ci, 97%+ coverage. Telemetry-privacy tests + deterministic calibration test.

[JSONbored]

Shipped in PR #740 (merged 06f0a1fc). Measurement-only outcome calibration: a maintainer-gated, read-only GET /v1/repos/:owner/:repo/outcome-calibration that reports (1) whether the deterministic slop score is predictive (do higher slop bands merge less often, over resolved PRs carrying a persisted band), and (2) the positive/negative/pending split of agent recommendations from the outcome ledger. It never auto-adjusts a live score — recalibration stays an owner-reviewed decision, like time-decay.