fix(tests): the mock embedding provider must not depend on the hash seed

[RobertSigmundsson]

Summary

• MockEmbeddingProvider derives its vector from the built-in hash(), which CPython salts per
  interpreter, so the mock returns different vectors from run to run.
• With a modulus of 1000, two different texts collide roughly once in a thousand runs, and
  test_different_texts_differ then fails on two byte-identical vectors.
• blake2b gives the same digest in every process, so a collision between two texts is now a
  property of the texts rather than of the seed the job happened to draw.

Why

Test (Python 3.11) failed on a pull request whose diff touched unified_config.py and a
consolidation test, several directories away from embeddings:

tests/unit/test_embedding_provider.py:134: in test_different_texts_differ
    assert v1 != v2
E   assert [0.4987766778715924, …] != [0.4987766778715924, …]

A mock written for determinism should be deterministic; a failure whose cause is the hash seed
teaches contributors to re-run the job rather than read it, which is a habit worth avoiding.

Changes

• tests/unit/test_embedding_provider.py: MockEmbeddingProvider.embed derives h from
  hashlib.blake2b(text.encode("utf-8"), digest_size=8) instead of hash(text).
• The same file gains test_embedding_does_not_depend_on_the_hash_seed, which pins the vectors
  for "text a" and "text b". On a salted hash it fails under every seed but the one that
  produced the constants, so the old behaviour cannot return unnoticed.

Test plan

• PYTHONHASHSEED=99 pytest tests/unit/test_embedding_provider.py on main — 1 failed, 82 passed, the failure being test_different_texts_differ with two identical vectors.
  Seed 99 is simply one that collides; most seeds pass, which is the bug.
• The same command on this branch — 84 passed. Also green under seeds 0, 1, 7, 12345 and
  424242, each 84 passed.
• With the change reverted and both tests kept, test_embedding_does_not_depend_on_the_hash_seed
  fails under all four seeds tried, and seed 99 additionally fails test_different_texts_differ.
• ruff check src/ tests/ clean; ruff format --check src/ tests/ reports 739 files already
  formatted.
• mypy src/ --ignore-missing-imports — success, no issues found in 354 source files.
• pytest tests/ --timeout=120 -m "not stress" against a live SurrealDB v3.2.0 on a freshly
  created namespace, serial as the Integration (SurrealDB) job runs it — 7286 passed, 48
  skipped, 1 xfailed, which is main's 7285 measured in the same environment plus the one
  test added here.
• CHANGELOG.md untouched — left to the release entry.

Related issues

Reported in #226.

Verified by @RobertSigmundsson

[acidkill]

Reviewed the full diff. blake2b with pinned constants is the right fix — deterministic across processes, and the regression test fails under a salted hash() regardless of seed. Green CI.