convert: --xbits e8 — produce and load the fmt=6 E8/IQ3 container (#452 step 4)

[ZacharyZcR]

Step 4 of #452, as promised there: the container becomes producible and loadable. Stacks conceptually on #465 (kernel) and #458 (codec), both merged.

What

Converter — --xbits e8 on convert_fp8_to_int4.py: routed-expert projections are rotated (W@Q, block-diagonal FWHT) and packed with the iq3 codec into fmt=6 bytes (98B/256 weights, 3.0625 bpw). All-or-nothing across gate/up/down — gate and up share one rotated input row in the engine, so a mixed layout would buy two gather buffers for nothing. Non-expert tensors take the existing paths untouched.

Engine — the container loads and runs on the CPU tier:

• qt_resolve_fmt: fmt=6 keeps the uniform name + name.qs convention; .qs is a single-float tag (real scales live inside the 98-byte blocks), and ns==4 is the discriminator — every other format ships at least O floats of scales.
• activation rotation at the placement the quant.h rule prescribes: once per moe() call for gate/up (all routed experts of a layer share the row, ~1.4 ms/token on GLM dims vs ~11 ms if done per expert), per-expert for down (its input is the per-expert hidden).
• fmt=6 stays off CUDA (qt_cuda_upload rejects it, same as fmt=5) and off every whitelisted fast path (XEXP/idot/Metal gate on fmt==2) — it flows down the generic CPU path where matmul_e8 dispatches. The CUDA sibling rides cuda: grouped-int4 (fmt=4) support in the expert-group kernels — opens the GPU tier to g64 and E8 containers (#334) #451's GroupDesc in a follow-up.

Rotation signs are regenerated, never stored. Both sides draw the same xorshift64* stream (seed 417+n); the block tiling is one expression (b = rem & -rem: 6144 → 2048+4096, 1536 → 512+1024). No rotation data in the container, no way for the sides to drift without the oracle catching it.

Tests

• test_e8_kernel gains a rotation oracle: the fixture now carries a rotated-weights section (Python-side encode(rotate_rows(W)), raw x, Python's Q^T x, float64 reference). The C side reproduces the rotation to 1.2e-7 and the rotated matmul to 9.3e-7 worst-rel.
• test_iq3_pack.py gains the frozen-PRNG golden bits (the constants are the spec — drift decodes every rotated container to garbage), tiling cases, and the orthogonality property ((W@Q)(Q^T x) ≈ Wx).
• make check 114/114 locally.

Limits, stated plainly

• e8 needs the input dim ≡ 0 (mod 256); GLM (6144/1536) and OLMoE qualify, the tiny oracle (128/32) cannot — so the end-to-end container proof is the OLMoE quality sweep (step 2 of the plan in fmt=5: a deployable E8-lattice grouped container — int3 that beats shipped int4 quality at 25% fewer bytes (design + plan) #452), next. This PR's correctness evidence is the two-sided fixture oracle above.
• Perf on the CPU tier is the scalar reference kernel for now; SIMD and the CUDA sibling are follow-ups once the OLMoE sweep validates the quality numbers on a real container.

[ZacharyZcR]

CI note: the Engine/linux/macos jobs will stay red until #500 merges — dev currently tracks 12 stale test binaries (from #368's conflict resolution), so make skips rebuilding tests/test_e8_kernel and the runner hits EACCES before this PR's code is even exercised. Green everywhere else; re-run after #500 lands.