#3236 [Other] tests: three suite gaps a live PC showed (no engine, setup, serve, or docs change)

本文

### Problem and result

On a live PC (Windows 11, RTX 4090, a Strata server serving on the default port 8080) three suite gaps made the tools tests fail for reasons unrelated to the code under test:

- `tools/test_iq_pack.py` died at import on a checkout without llama.cpp's gguf-py (`_paths.gguf_py()` exits hard there). After the rebase onto current main, upstream's new `HadamardTests`/`NativeExpertsV5Tests` classes hit `NameError: name 'iq_pack' is not defined` (18 errors).
- `tools/test_strata_mcp.py`: the discovery scans were answered by the live server on 8080, not by the tests' own servers.
- six `tools/test_conversation_cache_*.py` files fail when run from the repo root (`python -m unittest`), as the suite is run, because `tools/` is not on `sys.path`.

After: the iq_pack file imports without gguf-py (`iq_pack` needs numpy + `gguf_reader` only, so it is imported unguarded), the GGUF-fixture classes (Compatibility, Conversion, SplitArtifact, NativeExpertsV5) skip with a named reason, the MCP tests see only their own servers, and the cache tests run from the repo root. No engine, setup, serve, or docs change.

### Related discussion

Closed in the 0.2.0-beta tracker reset; rebased on main and re-verified per the reset note.

### Approach

Guarded import + `setUpClass` skip in `test_iq_pack.py` (`split_model`'s `gate1` default moved to runtime so the module imports without gguf-py); `DEFAULT_PORT` patch in `test_strata_mcp.py`; `sys.path` insert in the six cache tests.

### Validation

Passed, rebased on 0.2.0-beta main (`e88ee39d`), Windows 11, Python 3.15:

- `python -m unittest tools.test_iq_pack` and `python tools/test_iq_pack.py`: 3 ran / 4 skipped, exit 0 (both start forms; the 4 skips are the GGUF-fixture classes on a checkout without llama.cpp).
- `tools.test_strata_mcp`: OK, exit 0.
- `tools.test_conversation_cache_disabled` / `_growth` / `_isolation`: 14 tests OK, exit 0.

Not run: the full tools suite in CI; no engine, setup, serve, or docs code touched.

### Limits

The skips are conditional on gguf-py being absent; with llama.cpp cloned the classes run as before. No behavior change for any shipped code.