Pass target context as ctx_other so Gemma4 E4B MTP can init

[dai-yamashita]

Fixes #91.

Why

MTP.init/2 already accepts a sidecar :draft_model (#86). That is enough for Qwen 3.8. It is not enough for Gemma4 E4B.

gemma4-assistant requires llama_context_params.ctx_other at draft context construction — shared KV and the target's token embeddings. llama.cpp throws if the pointer is missing:

Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)

Context.create/2 never exposed the field. The NIF starts from llama_context_default_params(), so params.ctx_other stayed nullptr. Qwen's constructor leaves cparams.ctx_other as nullptr, which is why the existing sidecar path stayed green.

Creating a draft context on the assistant GGUF without a target still fails. That is llama.cpp's contract.

What changed

Context.create/2 takes optional :ctx_other. The NIF copies the raw llama_context* into params.ctx_other and Fine-keeps the peer so the target cannot be collected while the draft still holds the pointer.

MTP.init/2 always passes the target as the draft's :ctx_other. Callers do not set it. Qwen ignores it; Gemma4 needs it.

{:ok, target} = LlamaCppEx.load_model("gemma-4-E4B-it-Q4_K_M.gguf",
                                      n_gpu_layers: -1, load_mtp: true)
{:ok, head}   = LlamaCppEx.load_model("mtp-gemma-4-E4B-it-Q8_0.gguf",
                                      n_gpu_layers: -1, load_mtp: true)

{:ok, mtp} = LlamaCppEx.MTP.init(target, draft_model: head, n_draft: 3)

Same :draft_model API as Qwen 3.8. The only new requirement is the peer pointer on the draft.

:mtp_sidecar gained a second fixture pair behind LLAMA_SMOKE_MTP_E4B_MODEL / LLAMA_SMOKE_MTP_E4B_DRAFT_MODEL. Unset vars skip that module, so --include mtp_sidecar with only the Qwen pair stays green.

Model.n_embd_out/1 docs now note that gemma4-assistant has a narrow n_embd and a wide n_embd_out. The existing hidden-width guard still compares n_embd_out.

No @version bump. The fork mix.exs is still 0.8.43.

Measured

Unsloth gemma-4-E4B-it-Q4_K_M.gguf (4_971_171_584 B) + mtp-gemma-4-E4B-it-Q8_0.gguf (98_653_248 B), M3 Ultra, Metal, 2026-09-12:

• Hex 0.8.48: MTP.init/2 fails with the ctx_other throw above.
• This commit: MTP.init/2 returns a session; generate is non-empty; timing_us.ckpt == 0 (E4B is dense, not hybrid).
• Isolated Context.create(head) without a target still fails. Expected.

I did not re-run the Qwen 3.8 sidecar matrix. The NIF default remains omitted → nullptr.

Deliberately not changed

• Metal teardown abort. After a successful generate the BEAM can still exit 134 at ggml_metal_device_free (GGML_ASSERT([rsets->data count] == 0)). That is teardown, not init. Explain the Metal teardown abort and disable residency sets in smoke runs #78 already documents it. Not this PR.
• --spec-default stays unreachable. Same as MTP: separate sidecar draft model, llama.cpp b10435 (Qwen 3.8) #86.
• Hex artefact / app vsn. This commit does not publish 0.8.49.

Verification

Default suite (no model) plus the new guards:

• Context.create/2 rejects a non-%Context{} :ctx_other
• explicit ctx_other: nil is the omit path
• %Context{} carries the peer resource it was built from

Optional E4B gate (skips when unset):

GGML_METAL_NO_RESIDENCY=1 LLAMA_BACKEND=metal LLAMA_SMOKE_MTP_E4B_MODEL=/path/to/gemma-4-E4B-it-Q4_K_M.gguf LLAMA_SMOKE_MTP_E4B_DRAFT_MODEL=/path/to/mtp-gemma-4-E4B-it-Q8_0.gguf   mix test --include mtp_sidecar

Qwen-only --include mtp_sidecar is unchanged: MTPSidecarTest still requires LLAMA_SMOKE_MTP_MODEL + LLAMA_SMOKE_MTP_DRAFT_MODEL.

Review notes

• The Fine keep on LlamaContext.ctx_other is load-bearing for Gemma4. Qwen drops the C pointer, but the Elixir peer field is still set so the resource graph is the same for both architectures.
• I did not bump @version. Happy to add a changelog/release-guide note if you want this in 0.8.49.

[nyo16]

Thanks @dai-yamashita — this is the right fix, and it matches upstream exactly: common_speculative_init_result sets cparams.ctx_other = ctx_tgt for every draft context, and llama-context.cpp only reads
it for GEMMA4_ASSISTANT / EAGLE3 / DFLASH, so the Qwen in-file and sidecar paths are untouched.