feat(experimentation): results aggregation query and payload builder

[gagantrivedi]

• I have read the Contributing Guide.
• I have added information to docs/ if required so people know about the feature. (deferred — internal; docs land with the results UI.)
• I have filled in the "Changes" section below.
• I have filled in the "How did you test this code" section below.

Changes

Contributes to the experimentation results scorecard (v0.2). The ClickHouse aggregation + pure payload builder that feeds the stats kernel (#7769). No model, endpoints or task — those, plus the ORM-coupled orchestrator (metric specs + per-environment expected shares), land in the next PR.

• Results query — one single-pass query: the v0.1 exposures CTE (first-exposure dedup, quarantine, half-open window) joined to post-exposure metric events, with per-metric conditional aggregation → per metric × variant sufficient statistics (n, sum, sum_squares), plus per-variant identity counts for SRM. Per-aggregation unit value: occurrence 0/1, count, sum, mean (zero when an identity has no matching events). No conversion window in v0.2 (window = exposure → as_of).
• build_results_summary — pure: runs compare_to_control per treatment and srm_p_value on the counts. Data floor (n ≥ 50/arm, ≥ 5 conversions/arm for occurrence) → inference null; SRM only above 100 total identities; chance-to-win flipped for lower-is-better metrics. Raw stats kept for every variant (durable record for future engines); means/status/names derived client-side.
• Dataclasses MetricSpec / MetricResult / ResultsSummary (reusing the kernel's VariantStats / Inference); asdict yields the wire shape directly.

How did you test this code?

• Query (faked client): row→VariantStats mapping, per-aggregation unit-value expressions, post-exposure join + quarantine exclusion + half-open window, params, no-metrics count-only path.
• Builder: per-treatment inference, identity/conversion floors, zero-control-mean and missing-control nulling, lower-is-better flip, SRM balanced/imbalanced/not-computable, exact wire shape via asdict.
• pytest tests/unit/experimentation/ — 284 passed; ruff + mypy strict clean.

Note

The query's ClickHouse-specific semantics (LEFT JOIN attribution, avgIf over zero matching rows, toFloat64OrZero) are asserted by SQL substring only — there is no ClickHouse in CI (parked). The mean aggregation guards avgIf with a count check; this is the least-covered path until ClickHouse-in-CI lands.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

3 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
docs	Ignored	Preview	Jun 15, 2026 11:39am
flagsmith-frontend-preview	Ignored	Preview	Jun 15, 2026 11:39am
flagsmith-frontend-staging	Ignored	Preview	Jun 15, 2026 11:39am

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.50%. Comparing base (32c8582) to head (bdfef40).

Additional details and impacted files

@@                       Coverage Diff                        @@
##           feat/experiment-stats-kernel    #7781      +/-   ##
================================================================
- Coverage                         98.57%   98.50%   -0.07%     
================================================================
  Files                              1462     1461       -1     
  Lines                             56550    56409     -141     
================================================================
- Hits                              55743    55567     -176     
- Misses                              807      842      +35     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.