test(recall): stop passing removed semantic_seeds into link expansion

[nicoloboschi]

Summary

Fixes the test-api (2/3) failure currently red on main and therefore on every PR whose shard includes this test:

FAILED tests/test_link_expansion_scoring.py::test_activation_preserves_additive_score_across_fact_types
  TypeError: LinkExpansionRetriever.retrieve() got an unexpected keyword argument 'semantic_seeds'

#2683 ("remove unused graph seed inputs") dropped semantic_seeds from retrieve() — Link Expansion deliberately chooses its own bounded seeds, so it doesn't inherit the semantic arm's independent limits and thresholds. The scoring regression test added by #2679 still passed the kwarg.

Fix

Per direction from the code owner: semantic_seeds should not be used anymore, so the test is updated rather than the signature restored.

Removing the kwarg alone isn't sufficient — retrieve() now derives seeds internally via _find_semantic_seeds(conn, ...), and the test's fake connection is a bare object() with no fetch. So the fix stubs that lookup to return the same seed the test previously injected:

async def fake_find_semantic_seeds(_conn, _embedding, _bank_id, fact_type, **_kwargs):
    return [RetrievalResult(id=f"seed-{fact_type}", text="seed", fact_type=fact_type)]

The test's subject — that graph merge order matches Link Expansion's additive per-type score — and every assertion are unchanged:

assert [result.id for result in combined] == ["a", "b"]
assert world_results[0].activation == pytest.approx(math.tanh(0.5) + 0.9 + 0.3)
assert experience_results[0].activation == pytest.approx(math.tanh(1.0) + 0.7)

Testing

• tests/test_link_expansion_scoring.py — 1 passed.
• ruff check / ruff format clean.

The skills/hindsight-docs/** hunks are generated output from unrelated docs/integration PRs that landed without regenerating the bundle; the pre-commit generator refuses to commit without them.