perf(db): R.4 — desc nulls last, swept and measured site by site

[Willi363363]

Fourth and last step of track R, rebased on staging and carrying its own
R.4 ⬜ → ✅ tick. Independent of #315, #316 and #317 in code — but it edits
the same register file, see Before merging.

09-query-debt.md asked for the sweep H.2 could not do in passing: "every
.desc() in an index, against every query that orders on it". Four indexes are
written DESC NULLS LAST; two of their queries already matched, found by H.2
and by K. The other two were probed one at a time.

Measured, site by site

Fifty thousand leaderboard entries over two thousand players, and twenty-five
thousand coin movements over two accounts. Median of seven explain (analyze)
runs each.

Site	Before	After
coins.ts the ledger read	3.3 ms, Bitmap Heap Scan + Sort	1.5 ms, Index Scan + Incremental Sort
account.ts the export	1.2 ms, Bitmap Heap Scan + Sort	0.5 ms, Index Scan, no sort at all
leaderboard.ts all-time board	26.5 ms	26.5 ms — left alone
daily-board.ts the day's board	6.3 ms	6.3 ms — left alone

Only the two that moved are changed. That is the discipline the register
asks for: a query fixed on the strength of another query's number is exactly the
guess it exists to refuse. daily-board.ts filters on game.daily_day and never
on the entry's mode, so the partial (mode, score desc, …) index cannot serve
it whatever the ordering says.

The ledger's sort becomes incremental rather than disappearing because it also
orders on seq, which no index carries beside created_at — the tie inside one
created_at still sorts.

The register's own number was wrong, and is corrected

09-query-debt.md claimed the all-time board went 45 ms → 26 ms by changing
one ordering. It does not reproduce. All four spellings — inner, outer, both,
neither — measured 26.5 to 27.1 ms on the same seed, with identical plans, and
identical again with enable_seqscan off.

The reason is G.7: the board is distinct on (user_id) now, so its ordering
starts with user_id, and no index starts there and continues with score.
The 19 ms was recovered by that deduplication, not by the ordering — and the row
was written before that shape existed. Corrected in place rather than deleted,
because a number that turned out to belong to something else is the thing this
register is for.

One spelling, not two

newestClockFirst is exported from coins.ts and used by exportAccount. An
order that has to match an index in two files is one source of truth or it is
the same defect coming back in whichever file is edited second.

Tests

Two cases in coins-volume.test.ts, both failing with the bare desc put
back — the plan one because the index is no longer reachable with seqscan and
bitmapscan off, the SQL one because the text is no longer written.

Checks

• turbo run test --force --concurrency=1: 10 of 10 tasks, 3 866 passed, 0 skipped.
• turbo run typecheck lint --force: 22 of 22.
• pnpm format:check: green.

Before merging

#316 and #317 also edit 09-query-debt.md, which is at its 200-line cap. On this
branch the file is 195; on top of those two it will be over. Whichever merges
last gets rebased and re-trimmed, and the register is due a split either way —
recorded rather than done here, since it is not this step.

🤖 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 11:57am 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