perf(db): R.2 — the home asks for four rounds, not for all of them

[Willi363363]

Second step of track R, rebased on staging now that #314 is merged and
carrying its own R.2 ⬜ → ✅ tick. Independent of #315.

selectGameHistory had no LIMIT. home.ts took four rows with
.slice(0, RECENT_ROUNDS), in Node, after every participation the account had
ever had crossed the wire — and no index serves both halves, since the predicate
is participant.user_id and the order is game.started_at on the joined
table.

The number the register asked for and did not have

09-query-debt.md recorded the shape on 2026-09-17 and said so plainly: "not
measured, because the machine that would produce it has four rounds in it".
Seeded — 2 000 rounds for one player among 22 000, four in five of them
finished — and measured from Node, median of 20 runs after 3 warm-ups:

before   median 8.76 ms   2000 rows   601.7 KiB   quicksort of 2000
after    median 4.17 ms      4 rows     1.2 KiB   top-N heapsort of 4

The plan is barely the point, and that is the honest finding. Both shapes
still hash-join a sequential scan; the limit buys a top-N heapsort instead of a
full one, and about four milliseconds. What it buys that matters is six
hundred kilobytes that stop crossing the wire on every home page load —
invisible on a loopback, the payload between a Vercel function and Neon, and the
only half that grows with the account.

The trap, which the register did not name

A limit alone would have been a regression. The home was also filtering, on
ended_at, after the slice. Four abandoned rounds at the top of a history fill
the budget, and a player who has played all week is shown an empty list.

So HistoryWindow carries limit and finishedOnly, both defaulting to
off: exportAccount passes no window and still gets every row, unfinished
included, which is what an export is.

finishedOnly reads game.ended_at, not participant.submitted_at — somebody
who walked out of a round that ran to the end still played it.

Tests, checked against the absence of the piece each covers

Four new cases in history.test.ts. Not assumed — each run with its own piece
removed:

Piece removed	Cases that fail
the ended_at is not null predicate	1 (the one that would have blanked the home)
the limit	1
game.ended_at swapped for participant.submitted_at	2

Checks

• turbo run test --force --concurrency=1: 10 of 10 tasks, 3 868 passed, 0 skipped, against a local Postgres and Redis.
• turbo run typecheck lint --force: 22 of 22.
• pnpm format:check: green.
• 09-query-debt.md loses the entry in this PR, and the file is back at exactly 200 lines.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DsMRxyowr2Ch5RwiAa3t9z

[vercel]

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

Project	Deployment	Actions	Updated
wikifake	Ready	Preview	Sep 18, 2026 11:49am UTC

[qodo-code-review]

ⓘ Qodo reviews are paused because your trial has ended. Ask your workspace admin to add credits to resume reviews. Manage billing