Quantize the glm_tiny fixture to fmt=6 and fmt=4 expert weights

[lineape]

What to build

A reproducible, scripted step that produces glm_tiny routed-expert weights in fmt=6 (E8/IQ3) and fmt=4 (int4-grouped) that expert_load actually reads. Reuse the existing converters — iq3_pack.py (fmt=6) and convert_fp8_to_int4.py (emits fmt=4/5/6) — rather than writing a new quantizer. This is what lets the parity gate exercise the rotation-bearing format (fmt=6) and a no-rotation control (fmt=4); the default f32 oracle runs without conversion but never touches the fmt=6 rotation, so it can't catch the bug.

Acceptance criteria

• glm_tiny routed-expert weights available in fmt=6 and fmt=4.
• expert_load resolves both formats directly (no .qs-absent runtime-quant fallback for these routed experts).
• Routed experts are quantized; the shared expert may stay f32 (it's coordinator-only, outside the parity scope).
• The conversion is a checked-in script/command, not a hand-made binary; documented so CI can regenerate it.

Environment: codegraph-indexed — run codegraph explore first; refs in /home/eli/dev/_reference-docs-and-repos/colibri.