Summary

Stacked on Luce-Org#237 (Luce-Org/lucebox-hub). Addresses @howard0su's review comments on dflash/src/common/backend_factory.h:

• Line 57 ("shall we use the unsloth single-file MTP-in-target GGUF?"): --mtp-gguf is now optional. When absent, gguf_contains_mtp_tensors() probes the primary GGUF for qwen35.nextn_predict_layers > 0 and selects MtpSource::Native automatically.
• Line 59 ("why not a enum?"): enum class MtpSource { None, Native, ExternalDrafter, Auto } replaces the string-keyed mtp_draft_source and the load-shape-implied-by-pointer mtp_gguf_path.

Changes

// Before
const char * mtp_gguf_path    = nullptr;      // sentinel for shape
const char * mtp_draft_source = nullptr;      // "chain" | "mtp_topk"

// After
enum class MtpSource { None, Native, ExternalDrafter, Auto };
MtpSource    mtp_source       = MtpSource::None;
const char * mtp_gguf_path    = nullptr;      // only for ExternalDrafter
bool         mtp_use_topk     = false;        // false=chain, true=mtp_topk

dflash_server CLI

• --mtp-source [none|native|external|auto] — new explicit flag
• --mtp-gguf PATH — now optional; only required for ExternalDrafter mode
• --mtp-gamma N alone (no --mtp-source) triggers Auto detection
• Old --mtp-draft-source flag emits a stderr warning and is ignored (migration aid)

Auto-detect

gguf_contains_mtp_tensors(path) opens the GGUF, checks for qwen35.nextn_predict_layers > 0 — the same field qwen35_mtp_loader.cpp requires. Pure metadata scan, no tensor allocation, no GPU touch.

Backward compat

• --mtp-gguf PATH without --mtp-source → inferred as ExternalDrafter (old behavior preserved)
• --mtp-gamma N without --mtp-source and without --mtp-gguf → Auto (probes primary GGUF)

Tests

test_common_mtp_orchestrator (20-test mock-based suite) all pass. Build clean.

Note on tensor naming

gguf_contains_mtp_tensors() uses the qwen35.nextn_predict_layers metadata key as the detection heuristic — same criterion as qwen35_mtp_loader.cpp (which fails early with "not an MTP variant" if this key is absent). This covers the unsloth single-file format. If a different MTP GGUF format emerges without this key, howard can advise and we'll add a tensor-name scan as a fallback.