fix(draft): embed Qwen3.6 SWA metadata in Q8 converter

[Amidwestnoob]

Summary

• add an opt-in --qwen36-swa profile to the Q8 draft converter
• embed the Qwen3.6 draft's 2048-token sliding-window metadata and [SWA, SWA, SWA, SWA, full] layer pattern
• preserve the current Qwen3.5 conversion output by default
• add converter/metadata regression coverage and update the drafter recipe

Why

The draft GGUF loader already consumes these metadata fields, but quantize_draft_q8.py does not emit them. Qwen3.6 Q8 draft artifacts therefore require the external DFLASH27B_DRAFT_SWA=2048 runtime override to select the correct attention layout.

In a no-override negative control, the existing keyless artifact loaded without SWA metadata and matched only 6 of 12 deterministic outputs from the correctly configured reference. Prose and long-context outputs diverged.

This change makes newly converted Qwen3.6 Q8 drafts self-describing. It does not claim a new speedup; it reproduces the existing correct override-based configuration without relying on process environment.

Validation

• python3 server/tests/test_quantize_draft_q8.py
  • 6 tests passed
  • covers default behavior, exact key names/types/values, CLI parsing, full converter CLI output, and GGUF round-trip
• cmake --build server/build --target dflash_server -j2
  • built successfully
• ctest --test-dir server/build --output-on-failure -R "server_unit" --no-tests=error
  • 1/1 passed
• git diff --check
  • clean

Real Q8 artifact validation without DFLASH27B_DRAFT_SWA:

• loader: SWA layers: 4/5 (window=2048)
• HTTP success: 12/12
• output parity with env-configured reference: 12/12
• acceptance-rate parity: 12/12 exact
• peak VRAM: 20,760 MiB on both runs
• mean decode: 113.85 vs 114.70 tok/s (-0.74%, within run-to-run noise)

The metadata-added validation artifact preserved all 58 tensor payloads byte-for-byte; only the two intended metadata fields were added.

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}