Include part fields in the KiCad category listing (one request per category instead of one per part)

[arlenarlenarlen]

What

getCategoryParts() returns only id/name/description, so KiCad's symbol chooser has to fetch every part individually when it enumerates the library. This inlines the record getKiCADPart() already builds, which lets KiCad skip those requests.

Why

SCH_IO_HTTP_LIB::EnumerateSymbolLib() back-fills each part from parts/{id}.json unless the category listing already carried a fields object — HTTP_LIB_CONNECTION::SelectAll() uses its presence to set detailsLoaded (KiCad ≥ 10.0.5, common/http_lib/http_lib_connection.cpp). Because the listing omits it, the skip never triggers.

On the 351-part / 14-category instance I measured, that's 366 requests per cold library open instead of 15. Most of each request is connection setup and round trip rather than server time — for that instance, ~105 ms TCP+TLS and ~50 ms RTT against ~60 ms of server work — so the request count dominates and a cold open takes tens of seconds.

Measured on the same database, same disk, cold cache both times:

	before	after
HTTP requests per cold open	366	15
per-part back-fill requests	351	0
server time, cold cache	2.39 s	0.21 s
server time, warm cache	1.96 s	0.09 s
total payload	331 KiB	302 KiB
slowest single listing	10 ms	45 ms (174 parts, 149 KiB)

The total payload gets smaller: 14 batched responses carry less overhead than 351 individual ones. The largest category costs 45 ms cold to assemble, and the result is cached with the same Part/Category/Footprint tags as before, so it is rebuilt exactly as often as it was.

How

By wiring up the $minimal argument getCategoryParts() already accepts and that KiCadApiController already exposes as ?minimal=1. Today that flag has no effect on the response — the cache closure captures only $category, so $minimal changes the cache key and nothing else. After this change:

• ?minimal=1 → the light id/name/description shape (what the endpoint returns for everyone today),
• default → the same record the per-part endpoint returns, which is a superset of it.

So it also gives users the control asked for in #1464, and keeps a cheap listing available for clients that only need names.

Tests

Three cases added to tests/Services/EDA/KiCadHelperTest.php, using the existing EDADataFixtures (category 1 has a KiCad symbol, so its parts are EDA-visible): the default listing carries fields plus symbolIdStr, ?minimal=1 still returns exactly id/name/description, and both listings describe the same parts in the same order.

Checks run locally

On PHP 8.4 + SQLite (a cell of the CI matrix), in a container with the extensions CI installs:

bin/phpunit (full suite)	1955 tests, 4852 assertions, 0 failures, 1 skipped
bin/phpunit --filter KiCadHelperTest	27 tests (24 existing + the 3 added), 69 assertions
composer phpstan (level 5)	no errors
bin/console lint:yaml config --parse-tags	66 files OK
bin/console lint:twig templates --env=prod	176 files OK
bin/console doctrine:schema:validate --skip-sync	mapping OK

Not run: the MySQL and PostgreSQL matrix cells, PHP 8.2/8.3/8.5, and Codecov upload. The change builds a PHP array and issues no new queries, so it should be database-agnostic, but I didn't verify that empirically.

Two notes on the tooling while I was in there, both pre-existing and unrelated to this PR:

• vendor/bin/ecs can't run: ecs.php uses the deprecated return function (ECSConfig $ecsConfig) format, and the installed Easy Coding Standard passes an ECSConfig where the closure's signature demands a ContainerConfigurator, so it fails while building its container before analysing anything. Happy to send a separate PR converting ecs.php to the fluent API if that's wanted. I matched the surrounding style in KiCadHelper.php by hand instead.
• CONTRIBUTING.md says phpstan runs at --level=2; the actual composer phpstan script and phpstan.dist.neon both say level 5. I used level 5.

Refs #1464 — this covers the listing side of that request (a light vs. full listing); it does not add per-parameter selection.