feat(engine): stage D2 - mistral sliding-window attention (#79)

[tlkahn]

Summary

Stage D2 enablement for MODEL_MISTRAL (issue #79 / epic #55): llama-shaped dense path plus uniform sliding-window attention on every layer. No forward.c changes - D0 F7 mask/trim already covers the mistral config shape.

• Support gate accepts windowed and null-window mistral; qwen3/llama still reject sliding window
• weights_expected_names llama-shaped set (21 untied / 20 tied)
• gen_tiny_ckpt mistral recipe (sliding_window=8, omit pattern when 0) + committed fixtures
• GPU smoke + degenerate-window + window-boundary incremental==full
• Parity table records v0.1-era checkpoint with real sliding_window=4096

Cycle 0 audit

Confirmed against HF mlx-community/Mistral-7B-Instruct-v0.1-4bit-mlx/config.json:

Field	Value
model_type	mistral
sliding_window	4096 (int; v0.2+ is null - do not use for window parity)
tie_word_embeddings	false
rope_theta	10000.0
rope_scaling	absent
weight names	llama set (no qk-norm)

Config / forward delta vs llama

Field	llama	mistral v0.1	D2 impact
has_qk_norm	false	false	same
attention_bias	false	false	same
has_sliding_window	false	true when int	allowed
sliding_window_pattern	0	absent (default 0)	all layers local
rope_theta	500000 (llama3)	10000	plain rope

R1 decisions

1. No forward.c changes unless RED proved a hole (none did).
2. Gate allows sliding window only for mistral via REJECT_ALLOW_SLIDING_WINDOW.
3. Both windowed and null-window mistral pass the gate.
4. Generator omits sliding_window_pattern when 0 (get_dim_int rejects 0); gemma3 still emits pattern=2.
5. Parity id = mlx-community/Mistral-7B-Instruct-v0.1-4bit-mlx. Short-context parity is family enablement; window-boundary correctness owned by tiny_mistral GPU tests (CI will not exceed 4096 tokens). Chat mode not forced preemptively (no dry-run evidence yet).
6. Out of scope: engine: hoist prefill sliding-window mask build out of the per-layer attention call #92 prefill mask hoist, engine: KV cache backing store grows unbounded under sliding-window attention (ring-buffer follow-up) #93 ring-buffer KV, gemma3 interleaved pattern (D4).

Test plan

• make test (43 passed)
• ./tests/test_emodel_gate (windowed + null pass; llama/qwen3 SW rejects held)
• ./tests/test_weights / ./tests/test_model_config (mistral expected names + tiny fixture)
• ./tests/test_forward_mistral_gpu (load/map/shape, short incremental, degenerate window, boundary incremental==full)
• make test-gpu (mistral + D0/kvcache/llama green; pre-existing test_weights_gpu deepseek GGUF assert unrelated)
• make test-parity-script mistral absent-dir skip green (wrapper-skip-absent-mistral)
• Manual: scripts/parity_family.sh mistral when checkpoint present

Closes #79

[tlkahn]

TODO: mistral real-checkpoint parity on 32 GB+ Mac

Manual item from the test plan is still open:

• scripts/parity_family.sh mistral with checkpoint present

Blocked on hardware, not on the D2 diff. The canonical id is mlx-community/Mistral-7B-Instruct-v0.1-4bit-mlx (~4.0 GB 4-bit). parity_temp0.sh cold-loads that checkpoint twice (oracle mlx_lm.server, then mlxd run). On the current 16 GB machine that double load is memory-bound and not a reliable gate.

Do this later on a 32 GB+ Apple Silicon Mac (local or cloud MacStadium / EC2 Mac / similar):

MLXD_MLX_SERVE_BIN=/path/to/scripts/mlx_lm_server_wrapper.sh \
MLXD_PARITY_CKPT_MISTRAL=$HOME/.cache/mlxd/models/mlx-community--Mistral-7B-Instruct-v0.1-4bit-mlx \
MLXD_PARITY_MAX_TOKENS=16 \
scripts/parity_family.sh mistral

Notes for whoever picks this up:

• v0.1 with sliding_window=4096 is required for windowed family enablement. v0.2+ often has sliding_window: null and does not replace this check.
• There is no smaller real HF mistral id that still exercises D2; window-boundary correctness stays with tiny_mistral / test_forward_mistral_gpu (already green).
• Short-context text parity only (prompt + max_tokens << 4096). If raw instruct loops, retry with MLXD_PARITY_MODE=chat before blaming the forward path.
• This TODO does not block merging D2 on the strength of the unit/GPU suite; it is residual real-quant confidence.