fix(search): keep mypy clean under sentence-transformers 6.0.0

[hey-august]

What this changes

The problem

sentence-transformers 6.0.0 was integrated on 2026-08-18 and broke the [TYPECHECK] check on every open PR in this repository.

The pin is unbounded (sentence-transformers>=2.2.0), so CI picked up the new major version. mypy and torch are unchanged. Nothing in the tree moved, signalwire/search/ last changed on Aug 13 and these same branches passed Test on Aug 17.

All four errors are in signalwire/search/:

• Two errors come from model.model_name = model_name, where mypy now resolves the model to a torch nn.Module subclass whose __setattr__ accepts only Tensor and Module.
• An additional two errors come from if not SentenceTransformer, which reads as a truthy-function error because the TYPE_CHECKING import binds the real class and is never None.

The solution, in this PR

The fix annotates those two locals as model: Any, matching the dict[str, Any] declarations sitting beside them, and makes the availability guards explicit is None checks. It adapts to 6.x rather than capping the dependency, since a cap would transitively hold back transformers.

Two tidier-looking alternatives fail. setattr(model, "model_name", ...) trips ruff B010 and fails the LINT gate. Moving availability onto a separate boolean breaks test_load_model_no_library, because the tests simulate the missing dependency by patching the module-level name to None, so the guard has to keep reading that name. The model_name attribute itself also has to stay: it looks like a dead write, but query_processor.py:243 reads it back through getattr for set_global_model.

Verified mypy-clean against both sentence-transformers 6.0.0 and 5.6.0, ruff check and format clean, and the full unit suite green (5772 passed, 100 skipped).

Checklist

• bash scripts/run-ci.sh passes locally
• New tests assert on content (not just "does not raise" / "is not None")
• New test functions are type-annotated (mypy covers tests/)

run-ci.sh: 31 gates pass locally, including TYPECHECK (the gate this PR fixes), TEST, LINT, FMT and SIGNATURES.

But five fail for reasons unrelated to this change:

• SEMVER-DIFF, SURFACE-NATIVE and ROUTE-COLLISION expect other repos, porting-sdk and the SDK, at paths that do not match my filesystem organization.
• DOC-AUDIT cascades from SURFACE-NATIVE
• DOC-SURFACE reports an identical 64.3% on origin/main with and without this commit.

Leaving the box un-checked.

No new tests. This is a typing-only change with no behavioral surface of its own, and the guards it rewrites are already covered: test_load_model_no_library exercises the SentenceTransformer is None path by patching the module-level name to None. That test is also what constrained the fix, since it rules out moving the availability check onto a separate flag.

Does this change public API?

• No
• Yes — naming it here so a maintainer can land the matching
  infrastructure change:

Changing something that goes on the wire?

No. The change is confined to signalwire/signalwire/search/, which loads local embedding models. Nothing here is emitted, no enum values or parameter names move, and no generated types are touched.