k3 presets: --ignore-turn-stop so evidence runs decode the full budget

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: cursoragent

inference_engine/bridge/manifest.py

@@ -75,6 +75,10 @@ def _harness_preset(name: str, description: str, mode_flag: str) -> Preset:
                 "--drafter-id", "${ENV:KAKEYA_MAC_DRAFTER_ID}",
                 "--f-theta-dir", "${ENV:KAKEYA_MAC_FTHETA_DIR}",
                 "--s5-exact-full-attn", mode_flag,
+                # Evidence runs decode the full budget: without this the
+                # Gemma-4 <turn|> stop caps decode at ~8 tokens and the
+                # report fails the SPEEDUP_DECODE_TOKENS gate rule.
+                "--ignore-turn-stop",
                 "--n-samples", "{n_samples}",
                 "--max-new-tokens", "{max_new_tokens}",
                 "--block-size", "{block_size}",

tests/inference_engine/bridge/test_manifest.py

@@ -231,6 +231,7 @@ def test_harness_preset_resolves_env_and_params():
     assert argv[0:2] == ["python3", "scripts/research/k3_integrated_niah_eval_mac.py"]
     assert HARNESS_ENV["KAKEYA_MAC_VERIFIER_PATH"] in argv
     assert "--fused-specdecode" in argv
+    assert "--ignore-turn-stop" in argv  # full decode budget (gate rule)
     assert argv[argv.index("--n-samples") + 1] == "7"
     assert argv[argv.index("--max-new-tokens") + 1] == "96"
     assert argv[argv.index("--block-size") + 1] == "8"