Add a parts collection: appliance models and recalls

[ralyodio]

Adds /c/parts to nichedb: the free, regularly-updated data underneath appliance parts.

What this is

There is no free catalogue of appliance parts. Part numbers, prices and what fits what is what PartSelect and RepairClinic sell, and nothing open replaces it. What is free is the layer underneath, and this adds both halves.

energystar-models walks the EPA's per-category certified product datasets on Socrata. Brand, model number, UPC and specification across 23 categories: refrigerators, freezers, washers, dryers, dishwashers, cooking, dehumidifiers, room AC, air cleaners, fans, water heaters, furnaces, boilers, heat pumps, thermostats, pool pumps and commercial equipment. Keyless, public domain, rewritten daily. A live run pulled 1,168 models across two categories in 0.7s.

cpsc-recalls walks the CPSC REST service. Every recall with the model numbers and barcodes affected, the hazard, and the remedy owed. Keyless, public domain, ~50/month, newest dated the previous day. A live run pulled 103.

They join on model number and UPC, which is the point: a model page can say the thing is recalled.

Two traps encoded rather than left to be rediscovered

The consolidated Model Index (8wj2-sec8) is the obvious dataset and the wrong one. 1,796,876 rows, every category in one table, and Socrata reports data_updated_at as today. Its newest record on 2026-09-25 was 2025-12-09 — nine months behind. The daily timestamp is a republish of unchanged content and nothing in the response says so. A test asserts no preset points at it.

The datasets disagree about the date column. The index uses date_certified; most categories use date_qualified or date_available_on_market. Socrata answers a $where on a column that does not exist with 200 and [], so a wrong column name is indistinguishable from a category nobody certified. Every preset names the column measured to carry data, and an empty first page on a first walk throws instead of quietly reporting zero. Each preset also records the row count and newest date it actually returned, the way the socratacrime presets do, so a dataset that stops can be told from a quiet month.

Also

.gitignore said node_modules/ with a trailing slash, which matches directories only. A git worktree's node_modules is a symlink, so it was not ignored and git add -A staged it. One character.

test/adapters.test.js keeps a hardcoded list of valid collection slugs; parts is added to it.

Checks

• 26 new tests covering both adapters, the presets, and the collection wiring
• bun run lint clean (0 errors)
• Both adapters smoke-tested against the live upstreams, output inspected
• Full suite green apart from the Redis and Postgres integration tests, which need services this dev box does not run. The first CI run confirmed this: with those services present, its only failure was the registry slug list, fixed in b319296.

Deliberately not included

EPREL, the EU registry, is the one free source with actual spare-parts data (Ecodesign law makes manufacturers register parts availability and costs). It needs a free API key via EU Login, and its endpoints 403 without one. I did not write an adapter against a response shape I have never seen. That is the natural next PR once the key exists.

🤖 Generated with Claude Code