merge-train: #150 + #154 + #148 + #149 onto current main (conflicts resolved, tests green)

[FluffyAIcode]

What this is

A verified integration branch that merges the four open Mac PRs — #150 (runner Python pin), #154 (launcher bash-3.2 fix), #148 (launcher 2048 + full preset), #149 (runaway guard + codegen presets) — onto current main (which already has #152/#153 streaming), with all conflicts resolved and tests green. Merge this one PR to land all four at once, or use it as the reference resolution for merging them individually.

Verified merge order

#150 → #154 → #148 → #149

• feat(mac-bridge): pin workload Python interpreter + import gate (Layers B/C) + reusable skill #150 — clean (only adds runner_python.py/run_preset.py/docs + launcher PYBIN; main's launcher was untouched by fix(mac-chat): stream tokens live so the CLI doesn't look frozen on long answers #152/fix(mac-chat): actually stop [stream] lines interleaving (the part #152 squash-dropped) #153).
• fix(mac-launcher): bash 3.2-safe empty-array expansion (fixes "EXTRA[@]: unbound variable") #154 — 1 conflict: run_kakeya_mac.sh cmd=(…) line (feat(mac-bridge): pin workload Python interpreter + import gate (Layers B/C) + reusable skill #150's $PYBIN vs fix(mac-launcher): bash 3.2-safe empty-array expansion (fixes "EXTRA[@]: unbound variable") #154's ${EXTRA[@]+…}). Resolution = union: cmd=( "$PYBIN" … "${args[@]}" ${EXTRA[@]+"${EXTRA[@]}"} ).
• feat(mac-launcher): long-answer-safe defaults + full-mode validation preset #148 — 2 conflicts: manifest.py + test_manifest.py (preset additions at the same anchor). Resolution = keep both presets (launcher-dryrun-bash32 + launcher-full); run_kakeya_mac.sh auto-merged (2048 default).
• fix(mlx-fused): runaway-loop guard for greedy markdown-marker collapse on code prompts #149 — the heavy one, 4 conflicts because main's fix(mac-chat): stream tokens live so the CLI doesn't look frozen on long answers #152/fix(mac-chat): actually stop [stream] lines interleaving (the part #152 squash-dropped) #153 streaming touches the same files:
  • fused_specdecode.py (×3 signatures): union both params → on_commit=…, stop_on_runaway=True. Loop bodies auto-merged so each loop has both _emit(on_commit, …) (streaming) and the _trailing_runaway_drop guard.
  • k3_integrated_niah_eval_mac.py (×4): union — keep --chat-stream-stdout (streaming) + --chat-scripted-file/--fused-no-loop-guard (fix(mlx-fused): runaway-loop guard for greedy markdown-marker collapse on code prompts #149) args; each generate call gets both on_commit=on_commit, stop_on_runaway=_guard.
  • manifest.py / test_manifest.py: keep all presets (chat-stream-probe + codegen-degen-probe + codegen-guard-validate + the launcher ones).

Testing (on the merged result)

• ✅ pytest tests/inference_engine/bridge/ tests/backends/mlx/test_fused_specdecode.py tests/inference_engine/bench/test_k3_report_gate.py → 112 passed
• ✅ bash -n scripts/run_kakeya_mac.sh; combined launcher dry-run (empty EXTRA) → --max-new-tokens 2048, $PYBIN, ${EXTRA[@]+…} all present
• ✅ Merged fused_specdecode.py loops carry both streaming and the runaway guard; k3_… has streaming + native-ref + guard + _guard defined

If you merge the four PRs individually instead, follow the order/resolutions above. After landing, close this integration PR (and the four originals will show as merged/redundant).