roadmap(pillar 4): foundations & learning loop

[JSONbored]

Parent roadmap: #525

Goal

Land the small cross-cutting foundations the rest of the wave depends on, and close the feedback loop so recommendations and slop scoring improve from real outcomes — the durable moat and the answer to "why not just prompt my own agent."

Children

• refactor(signals): shared public-safe redaction module #542 — shared public-safe redaction module
• feat(learning): calibrate slop & recommendations vs merge outcomes #543 — calibrate slop & recommendations vs merge outcomes
• miner-facing post-merge reward & outcome attribution #702 — miner-facing post-merge reward & outcome attribution
• implement upstream TIME_DECAY_* scoring dimension (deferred; owner-reviewed) #703 — implement upstream TIME_DECAY_* scoring dimension (deferred; owner-reviewed)

Done when

• A single shared public-safe redaction primitive is exported and reused by every new public surface.
• Slop assessments and recommendations are logged and calibrated against real merged-vs-closed-vs-stale outcomes.

Cross-cutting acceptance criteria

Inherits the wave criteria (#525): npm run test:ci, 97%+ coverage, public/private output boundary.