perf: stop the ollama health probe fetching /api/tags twice per poll

[coaxk]

Spotted from the ollama side: subarr-next was hitting /api/tags far more often than the poll interval alone explains. It is not one call per poll, it is two.

The defect

tags = await client.tags()          # call #1, already contains the model list
client.reset_vision_cache()         # clears the cache that exists to prevent this
vision_resolved = await client.resolve_vision_model()
    -> installed_models() -> self.tags()   # call #2, identical payload

reset_vision_cache()'s own docstring says it is for "Settings save + model-pull completion". The health probe is neither, so it invalidated the cache on every poll and the resolver re-fetched what the probe was already holding. Settings polls every 8s, so the cost per poll was 2x /api/tags plus 1x /api/version, indefinitely.

The fix

resolve_vision_model() takes an optional installed list; the probe passes the names it just derived from its own tags() call.

Behaviour is deliberately unchanged in both directions:

• the cache is still reset each poll, so a model pulled outside subarr is still picked up within one refresh
• any caller passing nothing fetches exactly as before, which is why the existing test_vision_capability.py suite needed no edits at all

Simply dropping the reset_vision_cache() call would have been a smaller diff but would have stopped noticing externally pulled models, so it was rejected.

Verification

The test pins the cost and the behaviour together, so neither can regress on its own: it asserts exactly one GET /api/tags and that the vision model still resolves.

Confirmed red before the fix, for the right reason:

AssertionError: expected 1 /api/tags per probe, got 2:
['/api/tags', '/api/tags', '/api/version']

• affected tests: 28 passed (new test plus the whole existing vision suite)
• full suite: 1985 passed, 6 skipped
• ruff check and ruff format --check clean