Epic: validated auto-loosening for auto-tune.ts's tuning advisor, gated on a clean backtest

[JSONbored]

Problem

src/review/auto-tune.ts's tuning advisor (computeTuningRecommendations) already computes what a loosening recommendation would be (e.g. lowering a confidence floor, raising a scope cap) — the code just refuses to ever apply one. OverridePayload's own doc comment states it plainly: "a loosening recommendation never carries a payload (autonomous loosening is the regression risk the loop exists to avoid)." #8082's Pareto-floor backtest is exactly the missing risk-measurement that comment names as the reason loosening was never allowed to be automatic. Once a proposed loosening change can be backtested against real historical ground truth and shown to regress nothing, the original justification for refusing it no longer straightforwardly applies.

Area

src/review/auto-tune.ts (computeTuningRecommendations, OverridePayload, the apply path) + #8082's calibration/backtest primitives.

Proposal

• Once Decide + implement whether a REGRESSED backtest verdict should ever gate a merge (Phase 2) #8105 has resolved (establishing real production comfort/precedent for backtest-gated automated actions on PRs), revisit whether the same backtest discipline should extend to auto-tune.ts's own loosening recommendations: a loosening recommendation could carry a real OverridePayload when — and only when — a backtest (calibration: pure BacktestCase corpus builder from RuleFiredEvent/HumanOverrideEvent pairs #8083-calibration: render a backtest score/comparison report as Markdown #8088's primitives) shows the proposed change clears the Pareto floor (calibration: Pareto-floor comparator between two BacktestScoreReports #8086) against real historical ground truth (calibration: wire linked_issue_scope_mismatch into the shared signal-tracking module (fired + reversal) #8101/calibration: wire recordRuleFired/recordHumanOverride for the remaining isConfiguredGateBlocker codes (excludes linked_issue_scope_mismatch, covered by #8101) #8104's fired+reversed data), with the held-out split (calibration: deterministic seeded held-out/visible split of the backtest corpus #8087) guarding against overfitting to visible incidents.
• This is a materially bigger trust step than Decide + implement whether a REGRESSED backtest verdict should ever gate a merge (Phase 2) #8105 (which is about holding/blocking one PR for review) — this would let the system change its OWN live configuration autonomously, even if only ever in a backtest-verified-safe direction. Decompose into real sub-issues only once this bigger step is explicitly approved; this issue is the tracked intent, not a build-ready spec.
• A concrete, narrow first candidate worth naming now (not committing to): LINKED_ISSUE_SATISFACTION_CONFIDENCE_FLOOR specifically — calibration: wire linked_issue_scope_mismatch into the shared signal-tracking module (fired + reversal) #8101 already gives it real ground truth, and it's a single scalar value, the simplest possible loosening target to backtest against.

Deliverables

Deliberately none yet — decision-and-design only until #8105 resolves and the bigger step is explicitly approved.

Resources

• Epic: verifiable pre-merge backtest for ORB/AMS deterministic rules #8082 (parent epic), Decide + implement whether a REGRESSED backtest verdict should ever gate a merge (Phase 2) #8105 (the precedent decision this is blocked on)
• src/review/auto-tune.ts (computeTuningRecommendations, OverridePayload — read the "never carries a payload" comment in full before starting)
• calibration: wire linked_issue_scope_mismatch into the shared signal-tracking module (fired + reversal) #8101 (a concrete first candidate's ground truth source)

Boundaries

No autonomous config change of any kind happens as part of this epic without an explicit, later, maintainer-approved design — this is tracking the DIRECTION, not committing to the mechanism yet.

maintainer-only — a significant trust/safety-critical architecture direction, not a build task for contributors.

[linear-code]

JSO-5600