feat(mac): gemma-4 interactive CLI chat on the Kakeya-for-Mac (MLX) engine

[FluffyAIcode]

What

Interactive CLI chat with gemma-4 on the Kakeya-for-Mac engine — the full verifier/proposer/f_θ bounded-memory architecture, on scripts/research/k3_integrated_niah_eval_mac.py --chat (reuses the validated fused_specdecode_generate_mlx_trim per-turn sequence; NIAH evidence loop untouched).

f_θ runs by DEFAULT in chat (--force-f-theta auto-enabled unless --all-mlx-drafter):

PYTHONPATH=.:sdks/python python3 scripts/research/k3_integrated_niah_eval_mac.py \
  --verifier-path "$KAKEYA_MAC_VERIFIER_PATH" --drafter-id "$KAKEYA_MAC_DRAFTER_ID" \
  --f-theta-dir "$KAKEYA_MAC_FTHETA_DIR" \
  --s5-exact-full-attn --fused-specdecode \
  --sink-size 4 --window-size 64 --block-size 4 --max-new-tokens 1024 --chat

config	flags	what runs
default (chat)	--chat (no --all-mlx-drafter)	verifier + proposer + f_θ (f_θ restores 25 sliding layers each turn)
fast	--all-mlx-drafter --cuda-trim	verifier + proposer + S5 bounded KV (f_θ bypassed — S5 free lunch)
AR fallback	scripts/chat_mlx_kakeya.py	verifier-only over S5 bounded cache

Verified on Mac M4 (git-bus bridge) — proposer AND f_θ both live, f_θ DEFAULT-on

Without any explicit --force-f-theta:

turn	gemma-4	f_θ	proposer	bounded KV
capital of France?	The capital of France is Paris.	f_theta_ran=TRUE (25 sliding layers)	blocks=2, accept_len=4.0	12.2 MB
three primary colors	The three primary colors are red, yellow, and blue.	f_theta_ran=TRUE (25 layers)	blocks=4, accept_len=3.5	18.4 MB

On gemma-4 the restored sliding K/V are recall-irrelevant (exact layers carry recall — S5 free lunch), yet f_θ executes, exercising the full pipeline. On full-attention models the same f_θ path is load-bearing.

mac_gemma4_kakeya_ftheta_chat.log

ADR

ADR 0015 gains a section "Mac (MLX) interactive engine — full pipeline, f_θ default-ON": documents the engine, the gemma-4 recall-irrelevant-but-runs caveat, the forensic timeline (f_θ S5-bypassed 2026-06-12 by b3a04d0 "Optimize MLX adaptive S5 native smoke path"; proposer blocks=0 silent bypass caught by 0a6fb19 evidence gate), and the measured result.

Changes

• scripts/research/k3_integrated_niah_eval_mac.py — --chat/--chat-scripted full-engine REPL; --force-f-theta (default-ON in chat) bypasses the S5 short-circuit so f_θ runs; _gen_turn branches over all-MLX + torch paths; reports f_theta_ran/f_theta_layers/blocks/mean_accept_len.
• scripts/chat_mlx_kakeya.py — lightweight verifier-only AR chat.
• inference_engine/bridge/manifest.py — presets mlx-kakeya-fused-chat-{smoke,ftheta}, mlx-kakeya-chat-smoke; MAX_NEW_TOKENS 512→2048.
• docs/adr/0015-…md — Mac interactive engine + f_θ default-on + forensics.
• tests/inference_engine/bridge/test_manifest.py — preset coverage (31 presets, 100%).

Testing

• ✅ pytest tests/inference_engine/bridge/test_manifest.py — 30 passed, 100% coverage.
• ✅ Mac M4: f_θ DEFAULT-on (f_theta_ran=TRUE, 25 sliding layers, no explicit flag), proposer blocks>0, correct answers, bounded KV.

_{To show artifacts inline, enable in settings.}