calibration: register slop + quality score knobs (report-only) in LOOSENABLE_KNOBS

[JSONbored]

Problem

Epic #8211 track D. Once slop/quality decisions leave labeled evidence, both floors become registry candidates — entering exactly the way every knob must: REPORT-ONLY, proposals with full evidence, apply refused until each knob's consumption plumbing ships as its own reviewed change (the #8159 → #8176 sequence, repeated).

Requirements

⚠️ Required pattern. Registry entries only + advisor surfacing — applyMode report_only, candidates/hard-minimums chosen with written rationale (these floors gate contributor-facing verdicts; bounds tighter than close-confidence's), split seeds unique per knob, sample floors at least close-confidence's 50/12 given score noisiness.

• Both entries + the report-only proposals path (already generic) picking them up; knobs endpoint lists them with applyMode.
• The flip-to-live issues get filed only AFTER proposals with real evidence exist — note this explicitly in the entries' comments.

Deliverables

• Registry entries + invariant tests pinning bounds/seeds
• Advisor + status surfacing verified by tests

Links & Resources

#8211 (epic), the capture sub-issue (blocked-by), src/services/loosening-knobs.ts

Boundaries

report_only — NO consumption plumbing in this issue.

maintainer-only — registry bounds are gate-adjacent policy.

[linear-code]

JSO-5708

[JSONbored]

Design finding before any entry ships (from #8223's now-merged capture semantics):

1. slop_gate_score is a CEILING knob — the gate blocks when slopRisk >= slopGateMinScore (normalized confidence = risk/100, shipped 0.60). Loosening means raising the threshold, and the registry's evaluators (evaluateKnobLoosening/evaluateKnobDrift + buildConfidenceThresholdClassifier) assume FLOOR semantics throughout (candidates below shipped, classifier flags below-value). Registering slop today would compute inverted comparisons — report-only or not, wrong math must not surface as evidence.
2. quality_gate_score has no global shipped default — qualityGateMinScore is per-repo nullable (the capture only fires when a repo configured it), and a registry entry requires the shipped constant the whole discipline anchors on.

Both resolve in #8225's exact mandate (direction-aware Pareto orientation for the tightening unification): the registry gains an orientation: floor | ceiling axis with per-orientation bounds and classifier mapping, and slop enters report-only THERE, with written bounds rationale. Quality stays out until a global default exists (tracked here). Re-scoping this issue accordingly rather than shipping inverted evidence — the sequencing cost is one issue-hop, the alternative is a misleading advisor surface.