fix(mac-chat): stream tokens live so the CLI doesn't look frozen on long answers

[FluffyAIcode]

Summary

Fixes the Mac chat CLI on long answers (e.g. code generation): it (1) looked frozen and (2) once streaming was added, the output format was mangled by interleaved progress lines.

Issue 1 — looked frozen (non-streaming)

The interactive REPL ran the entire generation before printing anything. On a long answer at ~5 tok/s that's minutes of silence. Fix: an on_commit streaming callback in all three fused generate loops (safe _emit wrapper) + live incremental decode in the REPL, so the answer builds on screen from the first token.

Issue 2 — mangled format (interleaving)

The per-block [stream] blk=.. t=..s progress lines went to stderr while the answer delta went to stdout; on a shared terminal they interleaved into the text (在计算机[stream] blk=1 tok=2 t=13.6s). Fix: emit the [stream] timing line only on the non-interactive (scripted/bridge) path; the interactive CLI streams only the clean answer delta. Added --chat-stream-stdout so a non-tty bridge run can capture the exact clean live format.

Validation (Mac M4)

• Streaming works / not a deadlock: first run showed 72 [stream] blocks emitted continuously (first tokens at t=8.4s, then ~1 block/0.4s) — steady generation.
• Format clean: re-run with --chat-stream-stdout captured the answer as continuous markdown with 0 [stream] lines in the output:
  gemma-4 [根据pow的机制…]> 在计算机科学中，实现 \pow(base, exp)` 函数通常有两种主要的机制：快速幂算法…`

mac_chat_format_fix.txt
mac_chat_streaming_fix.txt

Changes

• inference_engine/backends/mlx/fused_specdecode.py: on_commit + _emit in all 3 fused loops.
• scripts/research/k3_integrated_niah_eval_mac.py: live clean streaming in the REPL; [stream] timing only on the scripted path; --chat-stream-stdout flag.
• inference_engine/bridge/manifest.py: mlx-kakeya-chat-stream-probe preset.

Testing

• ✅ pytest tests/backends/mlx/test_fused_specdecode.py tests/inference_engine/bridge/test_manifest.py (47 passed)
• ✅ On-device probe: streaming steady + clean format (evidence above)

Independent of #148/#149/#150; based on main.

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