perf(d1): avoid catalog scans for admissions

[harlan-zw]

🔗 Linked issue

Follow-up to #620 and #622.

❓ Type of change

• 📖 Documentation
• 🐞 Bug fix
• 👌 Enhancement
• ✨ New feature
• 🧹 Chore
• ⚠️ Breaking change

📚 Description

Admissions scanned the resolved Skill catalog and counted Repository siblings before filtering and pagination. The query now starts from admissions and counts each returned Repository once after selecting the page.

Live D1 comparisons returned identical rows, order, sibling counts, and totals across all three boards. Populated pages read 94–98% fewer rows, including both queries. No indexes or migrations change.

🤖 AI disclosure: Harlan Agent Kit modified this description. My AI open-source policy.

[harlan-github-agent]

🤖 MERGED

Harlan Agent Kit posted this automated review. It is not Harlan's personal review or approval. AI open source policy. Last updated: 2026-10-11 11:17 UTC.

GitHub merged this pull request.

No material findings were recorded.

e9fa0cbe-9538-4da6-a0e3-82ab8d16e934

[harlan-zw]

Agent verification, using read-only queries against production D1:

The previous loader and the rewrite returned identical rows, order, descriptions, stars, sibling counts, and totals. Comparisons covered pages 1, 2, and 999 on each board. Both paths used the same indexable Skill set.

Board and page	Cards	Reads before	Reads after	D1 duration before	D1 duration after
Week, 1	3	22,074	428	71.4 ms	6.8 ms
Month, 1	50	22,544	1,293	76.1 ms	3.4 ms
Month, 2	14	22,544	659	75.3 ms	2.7 ms
All, 1	22	23,048	619	69.0 ms	1.8 ms

After totals include both statements on populated pages. Empty pages use one statement and preserved their full totals. Every statement reported zero rows written. Each timing is one execution, not a latency percentile. Background jobs remained active.

The first count-only rewrite still scanned about 21,400 Skill rows. Admission-first CROSS JOINs removed that scan. Live query plans showed an admission scan followed by existing Skill and Repository index lookups. Repository counts also use existing indexes.

The comparison called the actual exported loaders through a read-only D1 adapter. Raw SQL, parameters, metadata, and results are retained in the session scratchpad.

[harlan-github-agent]

Harlan Agent Kit posted this automated review. AI open source policy.

Automated Review: READY.

Read the Review results