feat(mac-launcher): long-answer-safe defaults + full-mode validation preset

[FluffyAIcode]

Summary

Updates the one-command Mac launcher scripts/run_kakeya_mac.sh now that the long-decode wrap fix (PR #146) is merged, and adds an on-device validation preset so the launcher's full pipeline is guarded end-to-end (not just the --fast path).

Why

Before PR #146, the full engine degenerated into a runaway repeat past the ~1024-token native-cache ring wrap, so the launcher's default budget (1024) sat right at the coherence cliff. PR #146 fixed that (single-token commits once the sliding RotatingKVCache wraps), so long answers are now coherent — the default can be generous.

Changes

• scripts/run_kakeya_mac.sh:
  • Default --max-new-tokens 1024 → 2048 (the wrap is no longer a coherence cliff; FULL mode just forgoes the spec-decode speedup past it).
  • Documented the long-answer-safe behavior (header + FULL-mode banner + help).
  • Verified: bash -n clean; --dry-run and --help correct; all flags valid against the current harness CLI.
• inference_engine/bridge/manifest.py: new mlx-kakeya-launcher-full preset — invokes the launcher in FULL mode (f_θ verifier+proposer+f_θ) on a long scripted answer (请详细解释POW的工作原理, 1300 tokens) that crosses the wrap, with validate_reports=True (§4 liveness + §2.4 quality gate). mlx-kakeya-launcher-smoke stays for fast wiring checks.
• Tests: allowlist + validate-reports sets updated; added test_mlx_kakeya_launcher_full_preset_runs_full_mode_past_wrap.

The FAST path (--fast → --cuda-trim, all-KVCache) is immune to the ring-wrap bug by construction; only the FULL path uses the hybrid cache that received the fix.

Validation (Mac M4, via bridge — mlx-kakeya-launcher-full, 1300 tokens)

The launcher's FULL pipeline ran end-to-end and passed the on-device gate:

• exit_code = 0, evidence_gate_exit_code = 0 → §4 liveness + §2.4 quality gate passed.
• f_theta_ran = True (25 sliding layers) → full verifier+proposer+f_θ pipeline executed.
• tokens = 1241 (crossed the ~1024 wrap), mean_accept_len = 1.526 (single-token past the wrap, as designed), 3.81 tok/s, resident KV 235.7 MB.
• Output coherent with a clean structured conclusion — no 由于由于 runaway.

run_kakeya_mac_full_validation.txt

Testing

• ✅ pytest tests/inference_engine/bridge/test_manifest.py (32 passed)
• ✅ bash -n scripts/run_kakeya_mac.sh; --dry-run shows FULL-mode argv with --max-new-tokens 2048
• ✅ On-device mlx-kakeya-launcher-full (Mac M4, 1300 tokens): gate passed, f_θ ran, coherent past the wrap (evidence above)

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