Manage specialist token spend: bound the proposal set, stop discarding it, and name a stop condition

docs/conf.py

@@ -85,4 +85,4 @@
 
 html_title = f"{project} {version_number} documentation"
 
-external_projects_current_project = "Hyperloom"
+external_projects_current_project = "Hyperloom"

src/hyperloom/agents/kernel/tools/forge_gemm_tuning.py

@@ -19,12 +19,6 @@
 from pathlib import Path
 from typing import Any
 
-# Sibling import: kernel-agent tools cannot rely on the ``hyperloom`` import root.
-sys.path.insert(0, str(Path(__file__).resolve().parent))
-from _io_utils import truthy  # noqa: E402
-
-sys.path.pop(0)
-
 
 def _load_input_json(path: str) -> dict[str, Any]:
     if not path:

src/hyperloom/inference_optimizer/tests/test_accuracy_gate_units.py

@@ -317,7 +317,6 @@ def test_crash_and_success_produce_same_fingerprint(self, tmp_path):
 
     def test_workload_change_changes_fingerprint(self, tmp_path):
         """A change to workload shape (ISL) changes the fingerprint."""
-        import yaml as _yaml
 
         cfg1 = tmp_path / "c1.yaml"
         cfg2 = tmp_path / "c2.yaml"

src/hyperloom/inference_optimizer/tests/test_coordinator_async_batch2_unit.py

@@ -1366,7 +1366,7 @@ async def test_compose_prompt_orchestration_all_advisory_blocks(
         assert token in out
 
 
-def test_research_scout_seed_block_keeps_all_rounds(coord: Coordinator) -> None:
+def test_research_scout_seed_block_keeps_findings_and_questions_only(coord: Coordinator) -> None:
     from hyperloom.orchestrator.knowledge import research_hints
 
     research_hints.append_hints(
@@ -1410,11 +1410,14 @@ def test_research_scout_seed_block_keeps_all_rounds(coord: Coordinator) -> None:
 
     assert "hint one" in block
     assert "hint two" in block
-    assert '"name": "first"' in block
-    assert '"name": "second"' in block
     assert "question one" in block
     assert "question two" in block
     assert "ignore-me" not in block
+    # Proposals moved to the shared untested-proposal queue, which also drops
+    # the ones already benched; rendering them here as well would double them.
+    assert "Untested executable proposals" not in block
+    assert '"name": "first"' not in block
+    assert '"name": "second"' not in block
 
 
 @pytest.mark.asyncio

src/hyperloom/inference_optimizer/tests/test_coverage_boost2_unit.py

@@ -6,7 +6,6 @@
 from __future__ import annotations
 
 import json
-from types import SimpleNamespace
 
 
 # --------------------------------------------------------------------------- #

src/hyperloom/inference_optimizer/tests/test_critic_verdict_map.py

@@ -2073,10 +2073,16 @@ def _build_specialist_prompt_text() -> str:
     return system_prompt + "\n" + user_prompt
 
 
-def test_specialist_prompt_renders_top_6_target():
+def test_specialist_prompt_renders_proposal_target_and_ceiling():
     text = _build_specialist_prompt_text()
-    assert "AT MOST **6** entries" in text
-    assert "top-6" in text
+    # Section 8 states both numbers; Section 1 repeats them so the target is
+    # visible before the specialist starts working, not only at exit time.
+    assert "**2 entries is the norm, 4 the hard cap.**" in text
+    assert "**2 proposals is the norm, 4 the hard" in text
+    # Both the padding and the keep-going pressures need a stated counterweight.
+    assert "``empty=true`` is better than one" in text
+    assert "stop once" in text and "not the only stop" in text
+    assert "a coin-flip proposal is worse than none" in text
     assert "reviews each surviving variant" in text
 
 

src/hyperloom/inference_optimizer/tests/test_custom_framework.py

@@ -13,7 +13,6 @@
 from __future__ import annotations
 
 import os
-from pathlib import Path
 
 import pytest
 import yaml

src/hyperloom/inference_optimizer/tests/test_grid_size_prompt_unit.py

@@ -0,0 +1,33 @@
+# SPDX-FileCopyrightText: 2026 Advanced Micro Devices, Inc.
+# SPDX-License-Identifier: MIT
+
+"""The grid author's two prompt surfaces agree on where a grid comes from and how big it gets."""
+
+from __future__ import annotations
+
+
+def _grid_prompt_surfaces() -> list[str]:
+    from hyperloom.orchestrator.prompts.prompt_builder import (
+        _idea_generation_lines,
+        _format_grid_injection_hint,
+    )
+
+    return [_format_grid_injection_hint("explore") or "", "\n".join(_idea_generation_lines())]
+
+
+def test_both_grid_surfaces_state_the_same_target_and_ceiling():
+    for surface in _grid_prompt_surfaces():
+        assert "4" in surface
+        assert "maximum 6" in surface
+    assert all("Untested proposals (current cycle)" in s for s in _grid_prompt_surfaces()[1:])
+
+
+def test_the_stale_proposal_set_wording_is_gone():
+    from pathlib import Path
+
+    from hyperloom.orchestrator.prompts import prompt_builder
+
+    stale = "proposal_set drives the next"
+    assert stale not in Path(prompt_builder.__file__).read_text(encoding="utf-8")
+    md = Path(prompt_builder.__file__).parent / "orchestration.md"
+    assert "specialist proposal_set" not in md.read_text(encoding="utf-8")

src/hyperloom/inference_optimizer/tests/test_per_domain_prompts.py

@@ -465,7 +465,6 @@ async def test_runner_does_not_log_generic_template_for_any_domain(tmp_path):
     PolicyDenied,
     PolicyGate,
     SPECIALIST_ACTION_NAME,
-    SPECIALIST_FROM_AGENT_PREFIX,
 )
 from hyperloom.orchestrator.bus.resource_lock import (
     KNOWN_LANES,