… paths

Adds unit tests exercising MtpChainRunner state machine (gamma propagation,
EOS termination, partial-accept rollback, n_gen termination), MtpOrchestrator
lifecycle (reset_chain ordering, set_initial_hidden plumbing, gamma
derivation), and Qwen36MtpModule input-validation guards (attach(nullptr),
out-of-range gamma, pre-attach calls, shutdown idempotency).

All new tests are mock-based and require no GGUF.

Both qwen35_backend and the test_dflash harness called the 4-arg overload of
Qwen36MtpModule::init() with backend.tensor_context() — the BACKBONE ggml_ctx.
That ctx contains only blk.0..63; MTP head tensors live in blk.64.* in the
separate MTP GGUF, so the loader could never resolve them and init failed
silently before chain decode.

Switched both call sites to the 3-arg overload init(gguf_path, target, err)
which calls load_gguf_tensor_context(mtp_gguf_path) internally to build the
correct context for tensor lookup.

Caught by E2E sweep on real Qwen3.6-27B Q4_K_M + MTP-Q4_K_M GGUFs.

… is qwen35)

The GGUF general.architecture field on both 'Qwen3.5-0.8B-Q4_K_M.gguf' and
'Qwen3.6-27B-MTP-Q4_K_M.gguf' is 'qwen35'. The 'Qwen3.6' label lives only in
general.name (model display name). The MTP module is a NextN extension of the
qwen35 architecture, not a separate arch.

Aligns naming with Qwen35Backend (the backbone) and the loader's arch check
in gguf_target_loader.cpp:271. Pure rename — no functional change.

Files renamed (git mv, blame preserved):
  src/qwen36/qwen36_mtp.{h,cpp}         → src/qwen35/qwen35_mtp.{h,cpp}
  src/qwen36/qwen36_mtp_graph.{h,cpp}   → src/qwen35/qwen35_mtp_graph.{h,cpp}
  src/qwen36/qwen36_mtp_loader.cpp      → src/qwen35/qwen35_mtp_loader.cpp
  dflash/src/qwen36/ directory removed

E2E re-verified on Qwen3.6-27B Q4_K_M + MTP-Q4_K_M (RTX 3090):
  AR (γ=0):   35.6 tok/s
  Chain γ=2: 57.8 tok/s = 1.62× over AR  (accept_rate=0.83)
  Chain γ=4: 50.9 tok/s = 1.43× over AR  (accept_rate=0.58)
Greedy token output byte-identical across all three modes.

Mechanical fixes from style review:
- qwen35_mtp_loader.cpp: rewrite header from stale 'PR 2 skeleton'
  narrative to current behavior
- mtp_chain_runner.cpp: remove dead if-block at the post-propose
  size check (comment-only body)
- Qwen35DaemonArgs::n renamed to Qwen35DaemonArgs::mtp_draft_source
  to match Qwen35Config::mtp_draft_source
- mtp_orchestrator namespace: drop unused 'common' layer
  (dflash27b::common::mtp -> dflash27b::mtp)
- mtp_orchestrator: dynamic_cast -> static_cast (flavor() already
  discriminates the concrete type)
- qwen35_mtp.cpp: drop redundant 'static' on TU-local helpers inside
  anonymous namespace
- qwen35_mtp.{h,cpp}: convert Phase-A/B roadmap comments to TODOs or
  remove (completed phases)
- qwen35_backend.h: label ensure_decode_cache/tensor_context as
  test-only (they remain public for harness access)
- dflash_target.h: forward-declare DDTree; move ddtree.h include to
  .cpp files that use it concretely
- mtp_chain_runner.h: clean stale PR-internal cross-reference
- qwen35_backend.h: terminology consistency (MTP module)
- mtp_orchestrator.cpp: comment that enable_hidden_seq_capture is
  no-op for non-MTP targets
- test_common_mtp_orchestrator.cpp: remove PR-review narrative
- qwen35_backend.cpp: add fflush after [mtp] loaded printf
- No functional change. E2E sweep re-verified greedy byte-identical
  output across AR / chain γ=2 / chain γ=4.

Adds --mtp-gguf, --mtp-gamma, --mtp-draft-source, --mtp-draft-topk to
the native C++ HTTP server (dflash/src/server/dflash_server). Threads
them through BackendArgs into Qwen35Config so the existing Qwen35Backend
MTP path (already wired) is reachable through the production HTTP
endpoint, not just the test_dflash harness.

Mirror of the existing flags in dflash/scripts/server.py.

Verified via HTTP /v1/chat/completions: AR baseline and chain γ=2
produce greedy-identical text under temperature=0 on Qwen3.6-27B
Q4_K_M + MTP-Q4_K_M. Also confirmed with unsloth single-file combined
GGUF (Qwen3.6-27B-Q4_K_M-mtp.gguf) passed as --mtp-gguf — identical
output, greedy equivalent.

F2 (P1) mtp_chain_runner.cpp: on partial accept (accept_n < g_actual)
running_hidden was set from the drafter-internal next_hidden produced
after the full rejected suffix.  For ExternalDrafter flavor, call
set_capture_row(accept_n) + consume_captured_hidden() to thread the
target-captured hidden at the committed boundary instead.  NativeHeads
(Qwen3.6) is unaffected — next_hidden is always empty for that flavor
and the new branch is not entered.

F1 (P2) mtp_orchestrator.cpp: warm_and_decode chain-runner failure path
(line 151) returned without emitting io.emit(-1), leaving the stream
without a termination sentinel unlike every other error return in the
function.  Add the missing io.emit(-1) before return.

Cubic review: "On runner failure, stream termination sentinel io.emit(-1)
is not emitted, unlike all other error paths in this function" (P2).
"External-drafter prev_hidden desynchronizes on partial accept: running_hidden
is updated unconditionally from the full draft chain even when suffix drafts
are rejected, so the next iteration threads a hidden state from a rejected
token path instead of the committed boundary" (P1).

Build: clean.  All 19 unit tests pass.  E2E HumanEval-10 AR-vs-MTP
same-text rate: 8/10 post-fix (unchanged from pre-fix; F2 not active on
Qwen3.6/NativeHeads path — he_01/he_02 divergences confirmed GPU noise).

T7 used a NativeHeads stub so it never exercised the IExternalDrafterMtp
partial-accept path that commit 74f708a fixed. Add StubExternalDrafter +
T20 asserting set_capture_row(accept_n) is called before
consume_captured_hidden() when accept_n < gamma_actual.

…--draft/--mtp-gguf conflict

Real-workload testing with agentic CLI clients (hermes) exposed two
regressions not caught by HumanEval bench or unit tests:

A. DFLASH27B_MTP_CTX default was 8192; any prompt longer than that
   would print '[qwen35_mtp] warm_head_kv: n_prompt=N exceeds
   head_kv capacity n_ctx=8192' as a warning, then SIGKILL the server
   during the next prefill operation (CUDA assertion / memory
   corruption from writing past the buffer). VRAM spiked from 5.8GB
   to 20GB then process died. Silent — no error logged.
   Fix: default MTP_CTX to backbone max_ctx + warm_head_kv returns
   false on overflow so the chain runner falls back to AR cleanly.

B. dflash_server accepted --draft and --mtp-gguf together; the
   harness 'start_dflash_native_server' passes both by default which
   produced inconsistent state. Fix: at startup, log a WARNING and
   clear bargs.draft_path when both are set (MTP wins).

Verified: server now stays alive when handed a 15K-token prompt
under default MTP_CTX, falls back to AR for that request, returns
response. --draft+--mtp-gguf invocation logs the conflict warning
and loads only MTP.

P1 - gguf_mmap.h open() idempotency: call release() first if already
  open; on failure paths, reset state to default before returning.
P1 - gguf_mmap.h Windows release() HANDLE leak: CloseHandle() before
  clearing handle_. (Fix by inspection — Linux build does not exercise.)
P1 - qwen35_mtp.cpp GPU warmup off-by-one: n_prompt > n_ctx -> >= n_ctx
  matching CPU warmup. With slot_start=1, exactly-n_ctx-length prompts
  would have written slot n_ctx (out of bounds). MTP now declines that
  edge case cleanly and the chain runner falls through to AR.
P2 - test_dflash.cpp: missing hidden_seq capture now aborts cleanly
  (all_prefill_hidden.clear() + break) instead of running warm_head_kv
  on undefined buffer.
P2 - T7 (chain-runner rollback) now asserts restore_kv_at_chain was
  called > 0 times on partial accept.
P3 - T6 (EOS termination) tightened: total_emitted == 1, not <= 10.

Build clean, 20/20 unit tests pass. HumanEval-10 re-bench: 8/10
byte-identical (within AR-vs-AR noise baseline). Throughput in this
run reflects a pre-existing unstaged flashprefill.cpp change (BSA
opt-in -> opt-out) unrelated to these four files.

Per @howard0su's review on Luce-Org#237 (lines 57, 59):
- 57: 'shall we use the unsloth single-file MTP-in-target GGUF?'
- 59: 'why not a enum?'

Replaces:
  const char * mtp_gguf_path = nullptr;
  const char * mtp_draft_source = nullptr;  // "chain" | "mtp_topk"
with:
  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

Adds gguf_contains_mtp_tensors() probe (keyed on qwen35.nextn_predict_layers
metadata) so --mtp-gguf becomes optional when the primary GGUF embeds
MTP tensors (unsloth single-file case).

Stacked on Luce-Org#237. dflash_server arg parsing updated to:
- --mtp-source [none|native|external|auto] (new explicit flag)
- --mtp-gguf PATH (now optional; only needed for ExternalDrafter)
- Old --mtp-draft-source string flag warns + ignored (migration aid)
- --mtp-gamma alone triggers Auto detection

All test_common_mtp_orchestrator tests still pass (mock-based,
unaffected by the config-surface change).