fix(stats): correct the trial/pricing experiment funnel report (#679)

## 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.

app/Console/Commands/SendExperimentFunnelReportCommand.php

@@ -4,7 +4,9 @@ namespace App\Console\Commands;
 
 use App\Features\SubscriptionExperiment;
 use App\Services\Discord\DiscordWebhook;
+use App\Services\Stats\BinomialProportion;
 use App\Services\Stats\ExperimentFunnelCollector;
+use App\Services\Stats\ProportionSignificance;
 use Carbon\CarbonImmutable;
 use Illuminate\Console\Command;
 
@@ -22,8 +24,10 @@ class SendExperimentFunnelReportCommand extends Command
         SubscriptionExperiment::PAY_NOW => 'pay_now',
     ];
 
-    public function __construct(private ExperimentFunnelCollector $collector)
+    public function __construct(
-    {
+        private ExperimentFunnelCollector $collector,
+        private ProportionSignificance $significance,
+    ) {
         parent::__construct();
     }
 
@@ -42,6 +46,10 @@ class SendExperimentFunnelReportCommand extends Command
             $this->line($line);
         }
 
+        foreach ($this->significanceLines($report) as $line) {
+            $this->line($line);
+        }
+
         if ($this->option('no-discord')) {
             $this->info('Skipped Discord (--no-discord).');
 
@@ -66,30 +74,103 @@ class SendExperimentFunnelReportCommand extends Command
     {
         $revenue = $report['revenueAvailable'];
         $currency = $report['currency'];
-        $lines = [sprintf('%-8s %5s %5s %5s %5s %6s %8s %8s %8s %8s', 'Variant', 'Assg', 'Actd', 'Card', 'Paid', 'A2P%', 'MRR', 'Cost', 'Burn', 'CM')];
+        $lines = [sprintf(
+            '%-8s %5s %5s %5s %5s %5s %6s %7s %7s %7s %7s %7s',
+            'Variant', 'Assg', 'Actd', 'Card', 'MatU', 'Conv', 'Conv%', 'ARPU', 'MRR', 'Cost', 'Burn', 'CM',
+        )];
 
         foreach (self::LABELS as $key => $label) {
             $row = $report['variants'][$key];
-            $mature = $row['activatedMature'] > 0;
+            $mature = $row['assignedMature'] > 0;
+            $showMoney = $revenue && $mature;
 
             $lines[] = sprintf(
-                '%-8s %5d %5d %5d %5d %6s %8s %8s %8s %8s',
+                '%-8s %5d %5d %5d %5d %5d %6s %7s %7s %7s %7s %7s',
                 $label,
                 $row['assigned'],
                 $row['activated'],
                 $row['subscribed'],
-                $row['activeMature'],
+                $row['assignedMature'],
-                $mature ? ((int) round($row['activationToPaidRate'] * 100)).'%' : 'pend',
+                $row['convertedMature'],
-                $revenue ? $this->money($row['mrrCents'], $currency) : '—',
+                $mature ? ((int) round($row['conversionRate'] * 100)).'%' : 'pend',
+                $showMoney && $row['arpuCents'] !== null ? $this->money($row['arpuCents'], $currency) : '—',
+                $showMoney ? $this->money($row['mrrCents'], $currency) : '—',
                 $mature ? $this->money($row['costCents'], $currency) : '—',
                 $mature ? $this->money($row['wastedCostCents'], $currency) : '—',
-                $revenue && $mature ? $this->money($row['contributionMarginCents'], $currency) : '—',
+                $showMoney ? $this->money($row['contributionMarginCents'], $currency) : '—',
             );
         }
 
         return $lines;
     }
 
+    /**
+     * Per-variant conversion-rate uncertainty (95% Wilson interval) plus the
+     * leader-vs-runner-up verdict from {@see ProportionSignificance} — a Fisher
+     * exact test and a Newcombe difference interval, Bonferroni-corrected — so
+     * "check significance before calling a winner" has the numbers behind it.
+     *
+     * @param  array{startedAt: ?CarbonImmutable, currency: string, revenueAvailable: bool, costPerConnectionCents: int, variants: array<string, array<string, mixed>>}  $report
+     * @return list<string>
+     */
+    private function significanceLines(array $report): array
+    {
+        $lines = ['', 'Significance (95% Wilson CI on Conv%, n = MatU):'];
+        $arms = [];
+
+        foreach (self::LABELS as $key => $label) {
+            $row = $report['variants'][$key];
+            $n = (int) $row['assignedMature'];
+            $k = (int) $row['convertedMature'];
+
+            if ($n <= 0) {
+                $lines[] = sprintf('  %-8s pend (n=0)', $label);
+
+                continue;
+            }
+
+            [$low, $high] = $this->significance->wilsonInterval($k, $n);
+            $lines[] = sprintf('  %-8s %6s  [%6s – %6s]  (n=%d)', $label, $this->percent($k / $n), $this->percent($low), $this->percent($high), $n);
+            $arms[] = new BinomialProportion($label, $k, $n);
+        }
+
+        if (count($arms) < 2) {
+            $lines[] = 'Not enough matured variants to compare yet.';
+
+            return $lines;
+        }
+
+        usort($arms, fn (BinomialProportion $a, BinomialProportion $b): int => $b->rate() <=> $a->rate());
+        [$leader, $runnerUp] = [$arms[0], $arms[1]];
+        $result = $this->significance->compare($leader, $runnerUp);
+
+        $lines[] = sprintf(
+            'Leader %s vs %s: Δ %+.1f pts (95%% CI %+.1f … %+.1f pts, Newcombe).',
+            $leader->label, $runnerUp->label,
+            ($leader->rate() - $runnerUp->rate()) * 100, $result['diffLow'] * 100, $result['diffHigh'] * 100,
+        );
+        $lines[] = sprintf(
+            'Fisher exact p=%.3f %s α=%.3f (Bonferroni×3) -> %s.%s',
+            $result['fisherP'], $result['significant'] ? '<' : '≥', $result['alpha'],
+            $result['significant'] ? 'significant' : 'not significant',
+            $result['significant'] ? '' : ' Keep running.',
+        );
+
+        if ($result['minExpectedCount'] < 5.0) {
+            $lines[] = sprintf(
+                '(Small sample: min expected conversions %.1f < 5, so the normal-approx z=%.2f overstates — exact test used.)',
+                $result['minExpectedCount'], $result['z'],
+            );
+        }
+
+        return $lines;
+    }
+
+    private function percent(float $rate): string
+    {
+        return number_format($rate * 100, 1).'%';
+    }
+
     private function money(int $cents, string $currency): string
     {
         $symbol = match (strtolower($currency)) {
@@ -118,17 +199,22 @@ class SendExperimentFunnelReportCommand extends Command
                     'value' => $report['startedAt']->format('D, d M Y').' · new signups split evenly into the three variants.',
                     'inline' => false,
                 ],
+                [
+                    'name' => '📊 Significance',
+                    'value' => "```\n".implode("\n", $this->significanceLines($report))."\n```",
+                    'inline' => false,
+                ],
                 [
                     'name' => 'Legend',
                     'value' => sprintf(
-                        'Assg = signups · Actd = activated (connected a bank or enabled AI = cost triggered) · Card = completed checkout (card on file) · Paid = live non-refunded subs (mature) · A2P%% = Paid ÷ activated (mature) · MRR = monthly run-rate of paid subs (yearly ÷ 12) · Cost = est. connection cost of the mature cohort (%s/connection) · Burn = connection cost of mature users who never converted (money lost) · CM = MRR − Cost · `pend`/`—` = no mature data yet.',
+                        'Assg = signups · Actd = activated (connected a bank or enabled AI = cost triggered) · Card = completed checkout (card on file) · MatU = matured assigned (cohort old enough to score for this variant) · Conv = matured users who ever converted (were charged, net of refund) — time-invariant, so it does not shrink as an older cohort has longer to churn · Conv%% = Conv ÷ MatU (always ≤100%%, comparable across variants) · ARPU = MRR ÷ MatU (revenue per matured user) · MRR = monthly run-rate of *currently* paying subs (yearly ÷ 12); Conv above MRR is churn · Cost = est. connection cost of MatU (%s/connection) · Burn = connection cost of matured users who never earned net revenue (connected a bank but never paid, or paid then refunded) · CM = MRR − Cost · `pend`/`—` = no matured data yet.',
                         $this->money($report['costPerConnectionCents'], $report['currency']),
                     ),
                     'inline' => false,
                 ],
                 [
                     'name' => '⚠️ How to read it',
-                    'value' => 'Each variant is gated by its own decision window (control 15d, reduced 7d, pay_now 3d), so pay_now matures first — compare only once all three have mature volume, and check significance before calling a winner. **CM (contribution margin) is the decision metric.** Assg/Actd/Card are lifetime counts; A2P%/Cost/Burn/CM cover only the mature cohort, so the raw Actd→Card→Paid funnel mixes cohorts (immature carded users can\'t be Paid yet) — read it for volume, compare variants on A2P%/CM. Burn is what the connect-and-leave leak costs. Cost is a flat per-connection estimate across all providers, not per-provider billing.',
+                    'value' => 'Each variant matures on its own decision window (control 15d, reduced 7d, pay_now 3d, +3d settle), so at any moment MatU differs a lot between variants (pay_now matures first). **Compare variants on Conv% and ARPU — normalized per matured user — not on the absolute MRR/Cost/Burn/CM totals, which scale with MatU and so mechanically favour whichever variant has matured more.** Assg/Actd/Card are lifetime counts; everything from MatU rightward covers the matured cohort only, so the raw Actd→Card→Conv funnel mixes cohorts (immature carded users can\'t have matured yet) — read it for volume. Conv counts anyone ever charged (net of refund), so it is not depressed for older cohorts the way a live-active snapshot would be. Per-user CM is sub-cent at current volume, so treat CM as directional context, not the decision. Check significance (sample size = MatU) before calling a winner. Cost is a flat per-connection estimate across all providers, not per-provider billing.',
                     'inline' => false,
                 ],
             ],

app/Services/Stats/BinomialProportion.php

@@ -0,0 +1,22 @@
+<?php
+
+namespace App\Services\Stats;
+
+/**
+ * One arm of a binomial experiment: a count of successes out of trials, with a
+ * label. Bundles the (successes, trials) pair that the significance methods
+ * would otherwise pass around as loose integers.
+ */
+final readonly class BinomialProportion
+{
+    public function __construct(
+        public string $label,
+        public int $successes,
+        public int $trials,
+    ) {}
+
+    public function rate(): float
+    {
+        return $this->trials > 0 ? (float) $this->successes / $this->trials : 0.0;
+    }
+}

app/Services/Stats/ExperimentFunnelCollector.php

@@ -6,9 +6,9 @@ use App\Features\SubscriptionExperiment;
 use App\Models\User;
 use Carbon\CarbonImmutable;
 use Illuminate\Support\Facades\Cache;
+use Illuminate\Support\Facades\Log;
 use Laravel\Cashier\Cashier;
 use Laravel\Cashier\Subscription;
-use Laravel\Pennant\Feature;
 
 class ExperimentFunnelCollector
 {
@@ -20,12 +20,13 @@ class ExperimentFunnelCollector
 
     /**
      * Per-variant funnel for the trial/pricing experiment. Users are attributed
-     * by the variant Pennant resolved for them — the same value the runtime
+     * by SubscriptionExperiment::bucket() — the deterministic crc32 split that is
-     * served at checkout/paywall — so the report can't drift from what users
+     * the single source of truth for assignment — over the in-window signups the
-     * actually experienced (including any QA override or a legacy bucket that
+     * query selects, so it matches the variant each user was served without being
-     * predates the experiment). "Net active" is a live, non-refunded
+     * perturbed by the force_variant rollout hook (which pins every user to the
-     * subscription — an exact, heuristic-free metric that is comparable across
+     * winner once decided) or by Pennant store drift. "Net active" is a live,
-     * variants once each cohort clears its own decision window.
+     * non-refunded subscription — an exact, heuristic-free metric that is
+     * comparable across variants once each cohort clears its own decision window.
      *
      * The funnel is assigned → activated → carded (subscribed) → net-paying:
      * "activated" = the user connected a bank or enabled AI, i.e. triggered the
@@ -61,7 +62,9 @@ class ExperimentFunnelCollector
      *         refunded: int,
      *         assignedMature: int,
      *         activatedMature: int,
+     *         convertedMature: int,
      *         activeMature: int,
+     *         conversionRate: ?float,
      *         netActiveRate: ?float,
      *         activationToPaidRate: ?float,
      *         mrrCents: int,
@@ -92,8 +95,13 @@ class ExperimentFunnelCollector
         $excluded = (array) config('ai_suggestions.report.excluded_emails', []);
         $windows = $this->decisionWindows();
         $monthlyEquiv = $this->monthlyEquivByPriceId();
+        $missingPrices = [];
 
         User::query()
+            // Soft-deleted accounts still count: they were assigned a variant and
+            // their bank connections incurred real cost, and deleting the account
+            // is itself an experiment outcome (the strongest "connect and leave").
+            ->withTrashed()
             ->where('users.created_at', '>=', $startedAt)
             ->when($excluded !== [], fn ($query) => $query->whereNotIn('email', $excluded))
             ->with(['subscriptions' => fn ($query) => $query->where('type', 'default')])
@@ -104,11 +112,16 @@ class ExperimentFunnelCollector
                 'bankingConnections as connection_count' => fn ($query) => $query->withTrashed(),
                 'aiConsents as ai_consent_count',
             ])
-            ->chunkById(500, function ($users) use (&$variants, $windows, $now, $monthlyEquiv, $costPerConnectionCents): void {
+            ->chunkById(500, function ($users) use (&$variants, &$missingPrices, $windows, $now, $monthlyEquiv, $costPerConnectionCents): void {
-                Feature::for($users)->load([SubscriptionExperiment::class]);
-
                 foreach ($users as $user) {
-                    $variant = Feature::for($user)->value(SubscriptionExperiment::class);
+                    // Attribute by the deterministic bucket (the single source of
+                    // truth in SubscriptionExperiment), not the resolved Pennant
+                    // value: the latter is short-circuited by the force_variant
+                    // rollout hook, which would collapse every user onto one
+                    // variant once a winner is pinned. Every queried user is
+                    // in-window, so bucket() equals the variant they were served,
+                    // and reading it avoids writing Pennant rows as a side effect.
+                    $variant = SubscriptionExperiment::bucket((string) $user->id);
 
                     if (! isset($variants[$variant])) {
                         continue;
@@ -130,6 +143,21 @@ class ExperimentFunnelCollector
                     $status = $subscription?->stripe_status;
                     $netActive = $status === 'active' && $subscription->refunded_at === null;
 
+                    // "Converted" is time-invariant: the user was ever charged and
+                    // not refunded — currently active, or churned after the trial.
+                    // Unlike $netActive (a live snapshot), it does not shrink as an
+                    // older cohort has more time to cancel, so it is comparable
+                    // across variants that matured at different times. Excludes
+                    // trial-only cancels (ended on/before the trial → never charged).
+                    $converted = $subscription !== null
+                        && $subscription->refunded_at === null
+                        && $status !== 'trialing'
+                        && (
+                            $subscription->trial_ends_at === null
+                            || $subscription->ends_at === null
+                            || $subscription->ends_at->greaterThan($subscription->trial_ends_at)
+                        );
+
                     if ($subscription !== null) {
                         $row['subscribed']++;
                         $row['trialing'] += $status === 'trialing' ? 1 : 0;
@@ -154,10 +182,25 @@ class ExperimentFunnelCollector
                             $row['activatedMature']++;
                         }
 
+                        if ($converted) {
+                            $row['convertedMature']++;
+                        }
+
                         if ($netActive) {
                             $row['activeMature']++;
-                            $row['mrrCents'] += (int) ($monthlyEquiv[$subscription->stripe_price] ?? 0);
+                            $priceId = (string) $subscription->stripe_price;
-                        } else {
+
+                            if ($monthlyEquiv !== [] && ! isset($monthlyEquiv[$priceId])) {
+                                $missingPrices[$priceId] = true;
+                            }
+
+                            $row['mrrCents'] += (int) ($monthlyEquiv[$priceId] ?? 0);
+                        } elseif ($subscription === null || $subscription->refunded_at !== null) {
+                            // Burn = connections of matured users who never earned
+                            // net revenue: connected a bank but never carded, or
+                            // paid and got refunded. A user who paid and later
+                            // churned (canceled, not refunded) did convert, so
+                            // their connection cost is not burn.
                             $row['wastedCostCents'] += $connectionCostCents;
                         }
                     }
@@ -166,12 +209,21 @@ class ExperimentFunnelCollector
                 }
             });
 
+        if ($missingPrices !== []) {
+            Log::warning('Experiment funnel: net-active subscriptions on prices absent from the monthly-equivalent map — their MRR is undercounted as 0.', [
+                'price_ids' => array_keys($missingPrices),
+            ]);
+        }
+
         foreach ($variants as $key => $row) {
+            $variants[$key]['conversionRate'] = $row['assignedMature'] > 0
+                ? (float) $row['convertedMature'] / $row['assignedMature']
+                : null;
             $variants[$key]['netActiveRate'] = $row['assignedMature'] > 0
-                ? $row['activeMature'] / $row['assignedMature']
+                ? (float) $row['activeMature'] / $row['assignedMature']
                 : null;
             $variants[$key]['activationToPaidRate'] = $row['activatedMature'] > 0
-                ? $row['activeMature'] / $row['activatedMature']
+                ? (float) $row['activeMature'] / $row['activatedMature']
                 : null;
             $variants[$key]['arpuCents'] = $row['assignedMature'] > 0
                 ? (int) round($row['mrrCents'] / $row['assignedMature'])
@@ -189,7 +241,7 @@ class ExperimentFunnelCollector
     }
 
     /**
-     * @return array{assigned: int, activated: int, subscribed: int, trialing: int, trialingCanceling: int, active: int, canceled: int, pastDue: int, refunded: int, assignedMature: int, activatedMature: int, activeMature: int, netActiveRate: ?float, activationToPaidRate: ?float, mrrCents: int, arpuCents: ?int, costCents: int, wastedCostCents: int, contributionMarginCents: int}
+     * @return array{assigned: int, activated: int, subscribed: int, trialing: int, trialingCanceling: int, active: int, canceled: int, pastDue: int, refunded: int, assignedMature: int, activatedMature: int, convertedMature: int, activeMature: int, conversionRate: ?float, netActiveRate: ?float, activationToPaidRate: ?float, mrrCents: int, arpuCents: ?int, costCents: int, wastedCostCents: int, contributionMarginCents: int}
      */
     private function emptyRow(): array
     {
@@ -205,7 +257,9 @@ class ExperimentFunnelCollector
             'refunded' => 0,
             'assignedMature' => 0,
             'activatedMature' => 0,
+            'convertedMature' => 0,
             'activeMature' => 0,
+            'conversionRate' => null,
             'netActiveRate' => null,
             'activationToPaidRate' => null,
             'mrrCents' => 0,
@@ -218,8 +272,13 @@ class ExperimentFunnelCollector
 
     /**
      * Monthly-equivalent amount (in cents) for each plan price id, from Stripe.
-     * Yearly prices are divided by 12. Cached for an hour; returns [] (revenue
+     * Yearly prices are divided by 12. Fetched by product so that archived,
-     * unavailable) if Stripe can't be reached, without caching the failure.
+     * rotated price ids (Stripe mints a new id and transfers the lookup key on
+     * any amount change) still resolve — otherwise subscriptions on an old id
+     * would silently contribute 0 to MRR. Falls back to the current lookup keys
+     * when no product is configured. Foreign-currency and one-off prices are
+     * skipped. Cached for an hour; returns [] (revenue unavailable) if Stripe
+     * can't be reached, without caching the failure.
      *
      * @return array<string, int>
      */
@@ -231,26 +290,40 @@ class ExperimentFunnelCollector
             return Cache::get($key);
         }
 
+        $productId = config('subscriptions.products.pro');
         $lookups = array_values(array_filter([
             config('subscriptions.plans.monthly.stripe_lookup_key'),
             config('subscriptions.plans.yearly.stripe_lookup_key'),
         ]));
 
-        if ($lookups === []) {
+        if ($productId === null && $lookups === []) {
             return [];
         }
 
+        $params = $productId !== null
+            ? ['product' => $productId, 'limit' => 100]
+            : ['lookup_keys' => $lookups, 'limit' => 10];
+
         try {
-            $prices = Cashier::stripe()->prices->all(['lookup_keys' => $lookups, 'limit' => 10]);
+            $prices = Cashier::stripe()->prices->all($params);
         } catch (\Throwable) {
             return [];
         }
 
+        $currency = strtolower((string) config('cashier.currency', 'eur'));
         $map = [];
         foreach ($prices->data as $price) {
+            if ($price->recurring === null) {
+                continue;
+            }
+
+            if (strtolower((string) ($price->currency ?? $currency)) !== $currency) {
+                continue;
+            }
+
             $amount = (int) ($price->unit_amount ?? 0);
             $map[$price->id] = ($price->recurring->interval ?? 'month') === 'year'
-                ? intdiv($amount, 12)
+                ? (int) round($amount / 12)
                 : $amount;
         }
 

app/Services/Stats/ProportionSignificance.php

@@ -0,0 +1,149 @@
+<?php
+
+namespace App\Services\Stats;
+
+/**
+ * Frequentist inference on binomial proportions for the experiment funnel:
+ * per-arm Wilson intervals, a Newcombe interval for the difference of two
+ * proportions, and a Fisher exact test for the leader-vs-runner-up comparison.
+ *
+ * Fisher (exact at any sample size) is the decision test because the report's
+ * conversion counts are too small for the two-proportion z normal approximation
+ * to be valid — its expected cell counts fall well below the np>=5 rule.
+ */
+final class ProportionSignificance
+{
+    /** Two-sided 95% standard-normal quantile. */
+    private const Z_95 = 1.96;
+
+    /** @var list<float> memoised log-factorials, index i = log(i!) */
+    private array $logFactorials = [0.0, 0.0];
+
+    /**
+     * Compare the two leading arms: Newcombe difference interval, Fisher exact
+     * p-value, and a family-wise (Bonferroni) corrected significance flag. `z`
+     * and `minExpectedCount` are returned for the small-sample caveat only.
+     *
+     * @return array{alpha: float, diffLow: float, diffHigh: float, fisherP: float, significant: bool, minExpectedCount: float, z: float}
+     */
+    public function compare(BinomialProportion $leader, BinomialProportion $runnerUp, float $familyAlpha = 0.05, int $comparisons = 3): array
+    {
+        $alpha = $familyAlpha / $comparisons;
+        [$diffLow, $diffHigh] = $this->newcombeDiffInterval($leader, $runnerUp);
+        $fisherP = $this->fisherExactTwoSided(
+            $leader->successes, $leader->trials - $leader->successes,
+            $runnerUp->successes, $runnerUp->trials - $runnerUp->successes,
+        );
+
+        $pooled = ($leader->successes + $runnerUp->successes) / ($leader->trials + $runnerUp->trials);
+        $minExpectedCount = min($leader->trials, $runnerUp->trials) * min($pooled, 1 - $pooled);
+
+        return [
+            'alpha' => $alpha,
+            'diffLow' => $diffLow,
+            'diffHigh' => $diffHigh,
+            'fisherP' => $fisherP,
+            'significant' => $fisherP < $alpha,
+            'minExpectedCount' => $minExpectedCount,
+            'z' => $this->twoProportionZ($leader, $runnerUp),
+        ];
+    }
+
+    /**
+     * Wilson score interval for a binomial proportion — accurate for small n
+     * and near 0/1, where the normal approximation misbehaves.
+     *
+     * @return array{0: float, 1: float} lower and upper bound, clamped to [0, 1]
+     */
+    public function wilsonInterval(int $successes, int $trials, float $z = self::Z_95): array
+    {
+        $p = $successes / $trials;
+        $z2 = $z * $z;
+        $denom = 1 + $z2 / $trials;
+        $center = ($p + $z2 / (2 * $trials)) / $denom;
+        $margin = ($z / $denom) * sqrt($p * (1 - $p) / $trials + $z2 / (4 * $trials * $trials));
+
+        return [max(0.0, $center - $margin), min(1.0, $center + $margin)];
+    }
+
+    /**
+     * Newcombe (Wilson-based) 95% interval for the difference pA − pB. The
+     * correct object for "is A better than B": overlapping marginal intervals do
+     * NOT imply the difference includes 0.
+     *
+     * @return array{0: float, 1: float}
+     */
+    public function newcombeDiffInterval(BinomialProportion $a, BinomialProportion $b, float $z = self::Z_95): array
+    {
+        $pA = $a->rate();
+        $pB = $b->rate();
+        [$lA, $uA] = $this->wilsonInterval($a->successes, $a->trials, $z);
+        [$lB, $uB] = $this->wilsonInterval($b->successes, $b->trials, $z);
+
+        $lower = ($pA - $pB) - sqrt(($pA - $lA) ** 2 + ($uB - $pB) ** 2);
+        $upper = ($pA - $pB) + sqrt(($uA - $pA) ** 2 + ($pB - $lB) ** 2);
+
+        return [$lower, $upper];
+    }
+
+    /** Pooled two-proportion z statistic — descriptive only, not the decision test. */
+    public function twoProportionZ(BinomialProportion $a, BinomialProportion $b): float
+    {
+        $pooled = ($a->successes + $b->successes) / ($a->trials + $b->trials);
+        $se = sqrt($pooled * (1 - $pooled) * (1 / $a->trials + 1 / $b->trials));
+
+        return $se > 0.0 ? ($a->rate() - $b->rate()) / $se : 0.0;
+    }
+
+    /**
+     * Two-sided Fisher exact p-value for the 2x2 table [[a, b], [c, d]]
+     * (a/c = successes, b/d = failures). Sums the hypergeometric probabilities
+     * of every same-margin table no more likely than the observed one. Exact at
+     * any sample size — no normal approximation.
+     */
+    public function fisherExactTwoSided(int $a, int $b, int $c, int $d): float
+    {
+        $rowA = $a + $b;
+        $rowB = $c + $d;
+        $col = $a + $c;
+        $total = $rowA + $rowB;
+
+        if ($rowA === 0 || $rowB === 0 || $col === 0 || $col === $total) {
+            return 1.0;
+        }
+
+        $logProbObserved = $this->hypergeometricLogProb($a, $rowA, $rowB, $col);
+        $p = 0.0;
+        for ($x = max(0, $col - $rowB); $x <= min($col, $rowA); $x++) {
+            $logProb = $this->hypergeometricLogProb($x, $rowA, $rowB, $col);
+            if ($logProb <= $logProbObserved + 1e-7) {
+                $p += exp($logProb);
+            }
+        }
+
+        return min(1.0, $p);
+    }
+
+    private function hypergeometricLogProb(int $x, int $rowA, int $rowB, int $col): float
+    {
+        return $this->logChoose($rowA, $x) + $this->logChoose($rowB, $col - $x) - $this->logChoose($rowA + $rowB, $col);
+    }
+
+    private function logChoose(int $n, int $k): float
+    {
+        if ($k < 0 || $k > $n) {
+            return -INF;
+        }
+
+        return $this->logFactorial($n) - $this->logFactorial($k) - $this->logFactorial($n - $k);
+    }
+
+    private function logFactorial(int $n): float
+    {
+        for ($i = count($this->logFactorials); $i <= $n; $i++) {
+            $this->logFactorials[$i] = $this->logFactorials[$i - 1] + log($i);
+        }
+
+        return $this->logFactorials[$n];
+    }
+}

tests/Feature/SendExperimentFunnelReportCommandTest.php

@@ -7,8 +7,10 @@ use App\Models\User;
 use App\Services\Stats\ExperimentFunnelCollector;
 use Carbon\CarbonImmutable;
 use Illuminate\Support\Carbon;
+use Illuminate\Support\Facades\Artisan;
 use Illuminate\Support\Facades\Cache;
 use Illuminate\Support\Facades\Http;
+use Illuminate\Support\Facades\Log;
 use Illuminate\Support\Str;
 
 use function Pest\Laravel\artisan;
@@ -34,7 +36,7 @@ beforeEach(function () {
  * with an optional default subscription and any bank connections / AI consent
  * that mark the user as "activated" and drive the connection-cost columns.
  *
- * @param  array{status: string, at: CarbonImmutable, endsAt?: CarbonImmutable, refundedAt?: CarbonImmutable}|null  $subscription
+ * @param  array{status: string, at: CarbonImmutable, endsAt?: CarbonImmutable, trialEndsAt?: CarbonImmutable, refundedAt?: CarbonImmutable}|null  $subscription
  */
 function experimentUser(string $variant, CarbonImmutable $signup, ?array $subscription = null, int $connections = 0, bool $aiConsent = false): User
 {
@@ -52,6 +54,7 @@ function experimentUser(string $variant, CarbonImmutable $signup, ?array $subscr
             'stripe_price' => 'price_test',
             'created_at' => $subscription['at'],
             'ends_at' => $subscription['endsAt'] ?? null,
+            'trial_ends_at' => $subscription['trialEndsAt'] ?? null,
             'refunded_at' => $subscription['refundedAt'] ?? null,
         ]);
     }
@@ -173,6 +176,162 @@ it('computes connection cost, wasted burn and contribution margin', function ()
         ->and($control['activationToPaidRate'])->toBe(0.5);     // 1 paid ÷ 2 activated
 });
 
+it('keeps the A/B/C split even when a winner is forced, instead of collapsing onto one variant', function () {
+    // force_variant pins the runtime to one variant; the report must still show
+    // the real historical split, not attribute everyone to the forced winner.
+    config(['subscriptions.experiment.force_variant' => SubscriptionExperiment::PAY_NOW]);
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    experimentUser(SubscriptionExperiment::CONTROL, $signup);
+    experimentUser(SubscriptionExperiment::REDUCED_TRIAL, $signup);
+    experimentUser(SubscriptionExperiment::PAY_NOW, $signup);
+
+    $variants = app(ExperimentFunnelCollector::class)->collect()['variants'];
+
+    expect($variants[SubscriptionExperiment::CONTROL]['assigned'])->toBe(1)
+        ->and($variants[SubscriptionExperiment::REDUCED_TRIAL]['assigned'])->toBe(1)
+        ->and($variants[SubscriptionExperiment::PAY_NOW]['assigned'])->toBe(1);
+});
+
+it('warns instead of silently zeroing MRR when a paid sub is on an unmapped (rotated) price', function () {
+    Log::spy();
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    // The seeded map only knows 'price_test'; move this paid sub onto a rotated id.
+    $user = experimentUser(SubscriptionExperiment::REDUCED_TRIAL, $signup, ['status' => 'active', 'at' => $signup]);
+    $user->subscriptions()->first()->update(['stripe_price' => 'price_rotated_old']);
+
+    $reduced = app(ExperimentFunnelCollector::class)->collect()['variants'][SubscriptionExperiment::REDUCED_TRIAL];
+
+    expect($reduced['activeMature'])->toBe(1)
+        ->and($reduced['mrrCents'])->toBe(0); // unmapped price → 0, but now loudly
+
+    Log::shouldHaveReceived('warning')
+        ->withArgs(fn (string $message, array $context = []) => str_contains($message, 'absent from the monthly-equivalent map')
+            && in_array('price_rotated_old', $context['price_ids'] ?? [], true))
+        ->once();
+});
+
+it('still counts soft-deleted users so their assignment and connection cost survive', function () {
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    $user = experimentUser(SubscriptionExperiment::CONTROL, $signup, null, connections: 2);
+    $user->delete(); // soft delete the account
+
+    $control = app(ExperimentFunnelCollector::class)->collect()['variants'][SubscriptionExperiment::CONTROL];
+
+    expect($control['assigned'])->toBe(1)
+        ->and($control['assignedMature'])->toBe(1)
+        ->and($control['activated'])->toBe(1)     // 2 connections
+        ->and($control['costCents'])->toBe(80)    // 2 × 40, still charged
+        ->and($control['wastedCostCents'])->toBe(80); // never paid → pure burn
+});
+
+it('reports a matured-cohort conversion capped at 100% and prints the matured denominator', function () {
+    $signup = CarbonImmutable::parse('2026-06-05'); // control matures by the test clock
+
+    // Two paid, matured control users; only one is "activated" (connected a bank).
+    // The old A2P% (Paid ÷ activated = 2 ÷ 1) would print a nonsensical 200%.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, ['status' => 'active', 'at' => $signup->addDay()], connections: 1);
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, ['status' => 'active', 'at' => $signup->addDay()]);
+
+    $control = app(ExperimentFunnelCollector::class)->collect()['variants'][SubscriptionExperiment::CONTROL];
+
+    expect($control['assignedMature'])->toBe(2)
+        ->and($control['activatedMature'])->toBe(1)
+        ->and($control['activeMature'])->toBe(2)
+        ->and($control['convertedMature'])->toBe(2)
+        ->and($control['conversionRate'])->toBe(1.0); // Conv% = Conv ÷ MatU, always ≤ 100%
+
+    Artisan::call('stats:experiment-funnel', ['--no-discord' => true]);
+    $output = Artisan::output();
+
+    expect($output)->toContain('MatU')  // matured denominator column is printed
+        ->toContain('Conv%')
+        ->toContain('100%')             // capped, not the old 200% A2P%
+        ->not->toContain('200%');
+});
+
+it('reports a Wilson confidence interval and defers the verdict while samples are small', function () {
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, ['status' => 'active', 'at' => $signup->addDay()]);
+    experimentUser(SubscriptionExperiment::CONTROL, $signup);
+    experimentUser(SubscriptionExperiment::REDUCED_TRIAL, $signup, ['status' => 'active', 'at' => $signup->addDay()]);
+    experimentUser(SubscriptionExperiment::REDUCED_TRIAL, $signup);
+
+    Artisan::call('stats:experiment-funnel', ['--no-discord' => true]);
+    $output = Artisan::output();
+
+    expect($output)->toContain('Significance')
+        ->toContain('Wilson')
+        ->toContain('Fisher exact')       // verdict uses the exact test, not the z-approx
+        ->toContain('not significant')    // equal 50/50 rates, n=2 per arm → nowhere near
+        ->toContain('Small sample');      // min expected conversions < 5
+});
+
+it('declares significance via the exact test when the separation is real', function () {
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    // reduced: 5 matured converters; pay_now: 5 matured non-converters (no sub).
+    // Fisher exact on [[5,0],[0,5]] gives p≈0.008 < 0.0167 (Bonferroni) → significant.
+    for ($i = 0; $i < 5; $i++) {
+        experimentUser(SubscriptionExperiment::REDUCED_TRIAL, $signup, ['status' => 'active', 'at' => $signup->addDay()]);
+        experimentUser(SubscriptionExperiment::PAY_NOW, $signup);
+    }
+
+    Artisan::call('stats:experiment-funnel', ['--no-discord' => true]);
+    $output = Artisan::output();
+
+    expect($output)->toContain('Fisher exact')
+        ->not->toContain('not significant'); // the exact test clears the corrected bar
+});
+
+it('measures conversion as ever-charged, not active-now, so churn does not bias it', function () {
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    // 1) Still active → converted.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, ['status' => 'active', 'at' => $signup->addDay()]);
+    // 2) Paid then churned after the trial (charged, not refunded) → converted.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, [
+        'status' => 'canceled', 'at' => $signup, 'trialEndsAt' => $signup->addDays(15), 'endsAt' => $signup->addDays(40),
+    ]);
+    // 3) Canceled on/before the trial end (never charged) → NOT converted.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, [
+        'status' => 'canceled', 'at' => $signup, 'trialEndsAt' => $signup->addDays(15), 'endsAt' => $signup->addDays(10),
+    ]);
+    // 4) Refunded → NOT converted.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, [
+        'status' => 'canceled', 'at' => $signup, 'endsAt' => $signup->addDay(), 'refundedAt' => $signup->addDay(),
+    ]);
+
+    $control = app(ExperimentFunnelCollector::class)->collect()['variants'][SubscriptionExperiment::CONTROL];
+
+    expect($control['assignedMature'])->toBe(4)
+        ->and($control['convertedMature'])->toBe(2)   // #1 active + #2 churned-after-paying
+        ->and($control['activeMature'])->toBe(1)      // only #1 is active now
+        ->and($control['conversionRate'])->toBe(0.5); // 2 ÷ 4, time-invariant
+});
+
+it('excludes churned payers from burn but keeps refunds as burn', function () {
+    $signup = CarbonImmutable::parse('2026-06-05');
+
+    // Paid then canceled (not refunded): converted, so its cost is NOT burn.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, [
+        'status' => 'canceled', 'at' => $signup, 'endsAt' => $signup->addDays(20),
+    ], connections: 2);
+    // Paid then refunded: zero net revenue, so its cost IS burn.
+    experimentUser(SubscriptionExperiment::CONTROL, $signup, [
+        'status' => 'canceled', 'at' => $signup, 'endsAt' => $signup->addDay(), 'refundedAt' => $signup->addDay(),
+    ], connections: 3);
+
+    $control = app(ExperimentFunnelCollector::class)->collect()['variants'][SubscriptionExperiment::CONTROL];
+
+    expect($control['costCents'])->toBe(200)          // (2 + 3) × 40, all mature connections
+        ->and($control['wastedCostCents'])->toBe(120) // only the refunded user's 3 × 40
+        ->and($control['refunded'])->toBe(1);
+});
+
 it('scales connection cost by the cost-per-connection argument', function () {
     $signup = CarbonImmutable::parse('2026-06-05');
 

tests/Unit/ProportionSignificanceTest.php

@@ -0,0 +1,66 @@
+<?php
+
+use App\Services\Stats\BinomialProportion;
+use App\Services\Stats\ProportionSignificance;
+
+beforeEach(function () {
+    $this->stats = new ProportionSignificance;
+});
+
+it('computes the Wilson score interval', function () {
+    [$low, $high] = $this->stats->wilsonInterval(6, 52);
+    expect($low)->toEqualWithDelta(0.0540, 0.0005)
+        ->and($high)->toEqualWithDelta(0.2297, 0.0005);
+
+    [$low2, $high2] = $this->stats->wilsonInterval(5, 179);
+    expect($low2)->toEqualWithDelta(0.0120, 0.0005)
+        ->and($high2)->toEqualWithDelta(0.0640, 0.0005);
+});
+
+it('keeps the Wilson interval inside [0, 1] at the k=0 and k=n boundaries', function () {
+    expect($this->stats->wilsonInterval(0, 30)[0])->toBe(0.0)
+        ->and($this->stats->wilsonInterval(30, 30)[1])->toBe(1.0);
+});
+
+it('computes a two-sided Fisher exact p-value', function () {
+    // Real experiment table: reduced 6/52 vs pay_now 5/179.
+    expect($this->stats->fisherExactTwoSided(6, 46, 5, 174))->toEqualWithDelta(0.0182, 0.0005);
+    // Clean separation 5/5 vs 0/5 → 2 * C(5,5)/C(10,5) = 2/252.
+    expect($this->stats->fisherExactTwoSided(5, 0, 0, 5))->toEqualWithDelta(0.007936, 0.00001);
+    // Degenerate margin (no successes anywhere) → p = 1.
+    expect($this->stats->fisherExactTwoSided(0, 10, 0, 10))->toBe(1.0);
+});
+
+it('computes the Newcombe difference interval and the descriptive z', function () {
+    $reduced = new BinomialProportion('reduced', 6, 52);
+    $payNow = new BinomialProportion('pay_now', 5, 179);
+
+    [$low, $high] = $this->stats->newcombeDiffInterval($reduced, $payNow);
+    expect($low)->toEqualWithDelta(0.0164, 0.0005)
+        ->and($high)->toEqualWithDelta(0.2029, 0.0005)
+        ->and($this->stats->twoProportionZ($reduced, $payNow))->toEqualWithDelta(2.607, 0.01);
+});
+
+it('calls the borderline real comparison NOT significant under Bonferroni', function () {
+    $result = $this->stats->compare(
+        new BinomialProportion('reduced', 6, 52),
+        new BinomialProportion('pay_now', 5, 179),
+    );
+
+    // Fisher p ≈ 0.018 exceeds the corrected bar 0.05/3 ≈ 0.0167 → not a winner yet.
+    expect($result['significant'])->toBeFalse()
+        ->and($result['fisherP'])->toEqualWithDelta(0.0182, 0.0005)
+        ->and($result['alpha'])->toEqualWithDelta(0.0167, 0.0005)
+        ->and($result['minExpectedCount'])->toEqualWithDelta(2.48, 0.05);
+});
+
+it('calls a clean separation significant', function () {
+    $result = $this->stats->compare(
+        new BinomialProportion('a', 5, 5),
+        new BinomialProportion('b', 0, 5),
+    );
+
+    // Fisher p ≈ 0.008 clears the corrected bar.
+    expect($result['significant'])->toBeTrue()
+        ->and($result['fisherP'])->toEqualWithDelta(0.007936, 0.00001);
+});