test: make the MoE fixture actually test nesting, and fix two harness faults

[solderzzc]

Review follow-ups on #143/#144. All three findings were in code I wrote; the first is the substantive one.

1. The MoE fixture tested nothing it claimed

It used qwen3_moe, whose Swift configuration decodes from the root of config.json — so the expert count had to be at the top level, and the nested copy underneath it was never reached. findExpertCounts is breadth-first and returns on the first hit.

Both advertised properties were dead weight:

• the count being "nested under text_config rather than at top level"
• the vision_config: {num_experts: 999} decoy that "must not be mistaken"

Deleting the nested-container walk entirely would not have failed it.

Rebuilt on gemma4

Gemma4Configuration decodes text_config and nothing else, so the count exists only one level down and the decoy is genuinely reached. gemma4 also contains no "moe", so modelTypeImpliesMoE cannot rescue it — which is exactly what made the qwen3_moe version untestable.

Red-green verified — the check the previous version could not pass:

detection	result
pre-#114 (top-level num_local_experts only)	fails — 5 passed, 1 failed
current	passes — 6 passed, 0 failed

The fixture now reproduces #112.

The assertion also moved to the right gate

There are two gates, and I had been asserting on the wrong one. A config-level MoE check, and a model-level StreamableMoE conformance check. gemma4 passes the first and legitimately declines the second (it has no expert-streaming implementation), so asserting "streaming enabled" would have tested an unrelated capability. It now asserts only that detection did not reject — which is what #112 was about.

Incidentally this also covers the fused-expert remap: real gemma4 checkpoints ship experts.gate_up_proj as one tensor that sanitize splits into switch_glu.gate_proj/up_proj. A wrong split axis there is a silent numerical fault.

2. Port reuse could produce a false pass

cleanup() killed the server without waiting, and the readiness loop probed health before checking liveness. If a teardown outlived the 1 s sleep, the next fixture would fail to bind and its first probe would be answered by the previous server — assertions then run against the wrong checkpoint and report a false pass, not a flake.

cleanup now waits, and liveness is checked before the probe in both loops.

3. Regeneration was not idempotent

The generator wrote into existing directories without clearing them, so files a builder stopped emitting survived and the fixture kept testing a shape the source no longer described. Each fixture's own directory is now cleared first — scoped to one known directory, which is also what keeps regeneration away from siblings like tests/fixtures/omni, whose assets belong to test-omni.sh and are not generated here. Verified by planting a stale file, regenerating, and confirming it is gone while omni is untouched.

Verification

Suite green twice consecutively: 6 passed, 0 failed.

🤖 Generated with Claude Code