fix(stats): round global prCount at the per-year boundary

[detail-app]

Detail bug report: View on Detail

Bug

The unauthenticated statOutcomes (global) and statOutcomesByYear (per-year breakdown) endpoints both return the same StatOutcomes type with an Int prCount, but they rounded that field at different aggregation boundaries:

• Global: Math.round(Σ_all_years fractional_sum) — one Math.round on the aggregate.
• Per-year: Math.round(fractional_sum) per year-group, leaving callers to sum.

For pre-2025 years prCount uses the fractional estimator PRE_2025_PR_RATE = 0.71, so round(Σ f_i) ≠ Σ round(f_i) once pre-2025 events span ≥2 year-groups. Concrete case: two pre-2025 years with 1 matched project each → per-year gives round(0.71)+round(0.71)=2, global gave round(1.42)=1. The two views of the same field could disagree, with no in-repo consumer to notice. Introduced when the endpoints were added in a18bd13.

Fix

Round prCount at the same per-year-group boundary in both endpoints, so the global value is definitionally equal to the sum of the breakdown:

• Extracted the per-year GROUP BY query into a single shared projectsByYearQuery() used by both computeStatOutcomes and computeStatOutcomesByYear. The standalone global-aggregate prCount SQL is removed.
• The global prCount is now projectsByYear.reduce((acc, y) => acc + Math.round(y.prCount), 0) — the sum of the already-rounded per-year values.

Sharing the query structurally prevents the two rounding boundaries from drifting apart again, and the reduce (vs. the old prCount[0].prCount indexing) is robust to an empty GROUP BY result.

Testing

• Offline unit tests (src/resolvers/Stats.test.ts): added a regression guard, run with npx ts-node --transpile-only. It stubs Prisma and calls the private compute* methods directly to bypass the 1-hour ttlCache. It verifies the headline divergent case (two pre-2025 years → global prCount is 2, not 1), the empty/no-matched-project path returns 0 without crashing, that all StatOutcomes fields reconcile globally vs. the sum of the per-year view, and the structural invariant that both endpoints issue exactly one shared prCount-shaped query. Reverting the fix makes the suite fail; restoring makes it pass.
• Typecheck (npx tsc --skipLibCheck --noEmit) passes clean.
• Existing resolver test (src/resolvers/Review.test.ts) still passes — shared test infrastructure (env seeding, typedi Container, decorator metadata) is undisturbed.
• Lint: npx eslint on the changed files fails with a DeprecationError: 'originalKeywordKind' parser error, but this is a pre-existing, repo-wide incompatibility between the pinned eslint@7 and the installed typescript@5.2.2 — it reproduces identically on the unchanged Review.test.ts. No new source-level findings are introduced.
• End-to-end against live PostgreSQL (not versioned; temporary scaffolding, since removed): stood up PostgreSQL 16 in a Docker container and applied the schema via npx prisma db push --force-reset (prisma migrate deploy fails on a pre-existing, unrelated migration-ordering bug in the repo). Seeded Events for 2022, 2023, and 2025 with MATCHED projects (2025 with 3/5 carrying a prUrl), started a fresh GraphQL server process per variant, and issued a single unauthenticated query against both endpoints over HTTP. The populated seed returned statOutcomes.prCount = 5 reconciling with per-year 1+1+3 = 5 (the old aggregate behavior would have given round(4.42) = 4); an empty-seed variant returned prCount = 0 and statOutcomesByYear = [] without error. The resolver-level e2e was also confirmed to fail against the unfixed code and pass with the fix.

---

Automatic Fixes PRs can be configured here.

[augmentcode]

🤖 Augment PR Summary

Summary: This PR fixes an inconsistency between the global and yearly outcome statistics.

Changes:

• Extracts the matched-project aggregation into a shared projectsByYearQuery().
• Uses that grouped result in both statOutcomes and statOutcomesByYear.
• Changes global prCount to sum each year's individually rounded PR estimate.
• Preserves direct PR URL counting for 2025 onward and the 0.71 estimate for earlier years.
• Handles an empty grouped project result with a zero-initialized reduction.
• Adds offline resolver tests for fractional rounding, empty data, field reconciliation, and shared-query usage.

Technical Notes: The change makes the global prCount definitionally equal to the sum of the by-year values, avoiding differences caused by rounding before versus after aggregation.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. No suggestions at this time.

Comment augment review to trigger a new review at any time.