upstream: evaluate native FP8 dense projections and mixed-precision Qwen Flash loading #76

Separate loading correctness from the performance experiment of keeping dense projections quantized. Assessment baseline: fork 4fbc4ebf3f7d0c4039893b0471faf231c7a15ba0 (qualified RTX 4090 profile), upstream main af71ba4, donor moe-cache-drafting ae64bff9067c703730683a17e6e801908c6af096, inspected 2026-09-08 Pacific. Recheck current heads and existing PRs before implementation. Source inspection is not hardware qualification.

Candidate integration order:

1. FreeToken#404, 69b31ee09c674c443711793d973c03f82c758c22: model-specific mixed-precision expert/attention detection for NVIDIA Qwen Flash.
2. FreeToken#405, 06e4ec600150a8936147e0919b52c917f278da1a: resolve PLE from FTW source metadata. Reconcile with the fork's existing PLE discovery/streaming and relocation needs; a conversion-time absolute path must not become the only portable lookup.
3. FreeToken#392, 2b2a96eb53c8763d0ca8ec03a0f9cf2637281189: retain block-FP8 dense projections independently of NVFP4 experts; it stacks on #320. Isolate the incremental commit rather than importing duplicate loader ancestry.
4. Inspect #413 for llm-compressor NVFP4 export variants only when relevant to supported checkpoints.

The donor #392 reports 4.98 to 2.49 GiB dense storage and about 24.7% decode improvement on a PCIe-3 RTX PRO 4000 Blackwell. These are donor claims under a different model/runtime/hardware configuration, not projected 4090 gains.

Work: audit checkpoint declaration, per-module precision, fused GDN FP8 and BF16 projections, tensor scales, FTW serialization, expert-cache redistribution, and CPU/GPU backend capability. Preserve already-working RadixArk loading.

Acceptance: supported checkpoints load correctly, cold/restart PLE discovery works, native precision has a justified quality gate, and matched one/two-stream agent measurements show whether saved dense bytes/residency improve total wall. Record a no-benefit result if that is what the 4090 shows. Keep optional precision changes out of the qualified profile until validated.

Project relationships: Use #68 for measured qualification; residency alternatives are #69/#17 and #6.

Project: https://github.com/orgs/jomcgi-org/projects/2
Upstream and donor disposition ledger: #77