## Summary
Audited `stats:experiment-funnel` end-to-end (data acquisition, queries,
per-row counting, financial math, and statistics) and fixed the issues
that made it misleading for the win/no-win decision. The report used to
rank variants by an **absolute** contribution-margin total that
mechanically favoured whichever variant matured fastest, and it declared
statistical significance with a normal approximation that is invalid at
the small conversion counts this experiment has.
## What changed
**Presentation & comparability**
- Replace `A2P%` (numerator not a subset of the denominator → could
exceed 100%) with `Conv%` = conversions ÷ matured-assigned — always
≤100% and comparable across variants.
- Print `MatU` (matured cohort size) so the rate denominators are
visible and the table reconciles; revive `ARPU`.
- Reframe the guidance: compare on per-user `Conv%`/`ARPU`, not the
absolute `MRR`/`Cost`/`Burn`/`CM` totals (which scale with `MatU`).
**Data correctness**
- Count soft-deleted users (`withTrashed`) — they were assigned a
variant and their connections incurred real cost.
- Attribute by the deterministic `SubscriptionExperiment::bucket()`
instead of the resolved Pennant flag, so setting `force_variant` (the
winner rollout switch) no longer collapses the whole report onto one
variant. Also stops writing Pennant rows as a side effect.
- Resolve `MRR` for subscriptions on rotated/archived Stripe price ids
(fetch prices by product), warn on any net-active sub whose price is
unmapped, round yearly ÷ 12, and skip foreign-currency prices.
- `Burn` counts only users who never earned net revenue (no
subscription, or paid-then-refunded) — a paid-then-churned user is no
longer booked as connect-and-leave leak.
**Statistics**
- Measure conversion as "ever charged, net of refund" (time-invariant)
instead of a live active-now snapshot that biases older cohorts (which
have had longer to churn).
- Decide significance with **Fisher's exact test** (exact at any sample
size), Bonferroni-corrected over the three arms, instead of the
normal-approx z that overstates evidence when expected cell counts fall
below 5. Add a Newcombe difference-of-proportions CI and a small-sample
caveat.
- Extract the inference into `App\Services\Stats\ProportionSignificance`
(+ a `BinomialProportion` value object), with unit tests that pin the
exact interval and p-value numbers.
## Validation
Reconstructed every column in raw SQL against a production dump — all
reconcile **to the cent / to the row**. Independent recomputation of the
statistics (Wilson, Fisher exact, Newcombe, z) matches the command's
output.
## Testing
26 tests (20 feature + 6 unit, 94 assertions). `pint` and
`phpstan`/larastan green.
## Note
This branch also carries two small pre-existing commits unrelated to the
funnel (`fix(categories): fall back to gray…`, `chore(schedule): stop
scheduling the stuck cohort report`). Happy to split them into their own
PR if preferred.
3db03de86d