Improve test coverage of the hand-written core

[apdavison]

Coverage is 88% overall with a KG connection, 85% without. Both are flattered by generated code:

	statements	with token
generated (fairgraph/openminds/)	6647	91.1%
hand-written	1993	77.2%

The generated classes are constructor and property boilerplate, exercised incidentally by any test that touches them. The hand-written core is where the logic and the risk are.

Comparing a run with a token against one without separates two different problems.

Untested, no KG required — identical in both runs, so straightforwardly addressable:

module	coverage	statements missed
collection.py	21%	57
base.py	62%	22
caching.py	62%	10
embedded.py	82%	9
registry.py	88%	8

That is 106 statements reachable with plain unit tests, collection.py being far the largest single win.

Connection-gated — much better with a token, so the gap is CI's lack of credentials rather than missing tests: client.py 38%→67%, kgquery.py 44%→84%, kgproxy.py 59%→78%, node.py 85%→93%. Some of this can be pulled onto the offline path with a mock client, as was done for space_info() in #119; the rest belongs with running the live suite in CI.

(fields.py reports 0% but is a three-line deprecation shim due for removal at 1.0.)