feat(revenue_recovery): mathmodel retry-time predictor
Fills in the MathModel half of the Cascading retry strategy: computes when to
retry a failed subscription payment from a cluster's day-of-week / day-of-month /
hour-of-day success stats (dense-array `retry_stats` StatsDocument).
- `compute_mathmodel_retry_time(stats, budget, grace_days) -> Option<OffsetDateTime>`
in `workflows/revenue_recovery.rs`.
- Confidence-interval scoring (Laplace p̂, SE, pairwise Φ, logit-average), candidate-day
softmax + max combine, budget-aware probabilistic day/hour pick with a runway guard.
- Restricted exploration: n=0 (never attempted) and corrupt (k>n) slots are excluded
from scoring, so they can't hijack the pick.
- Window starts the NEXT day; `grace_days` counts the failure day, so the window is
`grace_days - 1` future days (capped at 31); returns `None` when `grace_days <= 1` (v1).
- Remove dangling `pub mod stub;` from `retry_stats.rs` (its file was deleted).
- Unit tests inline (`mod mathmodel_retry_time_tests`).
Not yet wired into the Cascading branch (min() with the static schedule) — follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
