feat(kernel): route mori EP dispatch/combine to the KernelForge vendor playbook

pyproject.toml

@@ -190,6 +190,11 @@ hyperloom = [
 "hyperloom.agents.kernel" = [
     "scripts/*.sh",
 ]
+# Vendor-operator-playbook registry (see tools/_vendor_operator_playbooks.py):
+# read at runtime by classify_patchability()/forge_submit.py, so it must ship.
+"hyperloom.agents.kernel.tools" = [
+    "vendor_operator_playbooks.json",
+]
 # ``agentx/deploy.py`` copies these into InferenceX ``benchmarks/`` at runtime
 # and raises FileNotFoundError when they are absent from the install.
 "hyperloom.inference_optimizer.assets.agentx" = [

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

@@ -0,0 +1,201 @@
+###############################################################################
+# SPDX-FileCopyrightText: 2026 Advanced Micro Devices, Inc.
+# SPDX-License-Identifier: MIT
+#
+# See LICENSE for license information.
+###############################################################################
+
+"""Vendor-operator-playbook registry: route a closed-source hot kernel to a
+validated KernelForge *task bundle* instead of a source rewrite.
+
+Most of the forge-submission pipeline assumes a hot kernel has an editable
+device source file (``kernel_url`` -> in-place rewrite). Some vendor
+operators -- mori's EP dispatch/combine all-to-all is the first case -- are
+pip-installed compiled libraries with no such source, but do have a small,
+named set of launch-config knobs that a KernelForge forge-loop task bundle
+has already been validated to tune (see KernelForge PR #88's "Making this
+real" section for the design rationale this module implements).
+
+This is deliberately a narrow, explicit carve-out (one JSON registry, sibling
+to the retired ``op_to_source.json``) rather than a general "config-tuning"
+system: a candidate only gets vendor-playbook treatment when it matches a
+registry entry by name.
+"""
+
+from __future__ import annotations
+
+import copy
+import functools
+import json
+import os
+from pathlib import Path
+from typing import Any
+
+_REGISTRY_PATH = Path(__file__).resolve().parent / "vendor_operator_playbooks.json"
+
+
+@functools.lru_cache(maxsize=1)
+def load_vendor_operator_playbooks() -> tuple[dict[str, Any], ...]:
+    """Load and cache the vendor-operator-playbook registry.
+
+    Returns:
+        A tuple of playbook entry dicts (empty when the registry file is
+        missing or malformed -- a missing registry must never be fatal to
+        the rest of the classification pipeline).
+    """
+    try:
+        raw = json.loads(_REGISTRY_PATH.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return ()
+    playbooks = raw.get("playbooks") if isinstance(raw, dict) else None
+    if not isinstance(playbooks, list):
+        return ()
+    return tuple(entry for entry in playbooks if isinstance(entry, dict) and entry.get("id"))
+
+
+def _reset_vendor_operator_playbooks_cache() -> None:
+    """Clear the cached registry (tests only, e.g. after monkeypatching the path)."""
+    load_vendor_operator_playbooks.cache_clear()
+
+
+def _candidate_haystack(candidate: dict[str, Any]) -> str:
+    """Join every text field a playbook's ``any_marker`` may match against."""
+    fields = (
+        candidate.get("name"),
+        candidate.get("operation"),
+        candidate.get("library"),
+        candidate.get("source_file"),
+        candidate.get("kernel_repo"),
+    )
+    return " ".join(str(f or "") for f in fields).lower()
+
+
+def _last_symbol_segment(value: str) -> str:
+    """Return the trailing method/function segment of a qualified symbol.
+
+    ``mori::EpDispatchCombineOp::combine`` -> ``combine``; a plain name with
+    no separator is returned unchanged. Needed because a class name like
+    ``EpDispatchCombineOp`` itself contains the substring "dispatch", so
+    matching a role marker against the *whole* qualified name is ambiguous --
+    only the actual called method disambiguates dispatch vs combine.
+    """
+    tail = value
+    for sep in ("::", ".", "/"):
+        tail = tail.rsplit(sep, 1)[-1]
+    return tail
+
+
+def _role_haystack(candidate: dict[str, Any]) -> str:
+    """Return the field(s) a playbook's ``name_any`` (op-role pattern) should match.
+
+    Prefers ``operation`` (often already the specific call, e.g.
+    ``"combine"``) over ``name`` (which may be a fully-qualified
+    ``Class::method`` symbol whose class name can itself contain another
+    role's marker); either way, only the trailing symbol segment is
+    matched, never the whole qualified string. This repo's own convention
+    (``_task_group_contract.logical_operator_name()``,
+    ``_bypass_report.py``'s task-group builder) is to set ``operation`` to
+    the fully-qualified name too, e.g. ``mori::EpDispatchCombineOp::combine``
+    -- taking ``operation`` verbatim would silently reintroduce the exact
+    dispatch/combine ambiguity this function exists to resolve the moment
+    some producer starts populating that field on a candidate row (PR #1191
+    review finding #6).
+    """
+    operation = str(candidate.get("operation") or "").strip()
+    if operation:
+        return _last_symbol_segment(operation).lower()
+    return _last_symbol_segment(str(candidate.get("name") or "")).lower()
+
+
+def match_vendor_operator_playbook(candidate: dict[str, Any]) -> dict[str, Any] | None:
+    """Return a matched playbook entry for ``candidate``, or ``None``.
+
+    A candidate matches a playbook when at least one of the playbook's
+    ``any_marker`` strings appears somewhere in the candidate's identifying
+    fields (name/operation/library/source_file/kernel_repo) AND at least one
+    of its ``name_any`` strings appears in the candidate's name/operation --
+    e.g. mori's playbook requires both "mori" (library/source evidence) and
+    "dispatch" or "combine" (which op within mori this is).
+
+    Args:
+        candidate: The hot-kernel candidate dict (as built by
+            ``tracelens_analysis``).
+
+    Returns:
+        A deep copy of the matched registry entry, augmented with a
+        ``"role"`` key set to whichever ``name_any`` marker matched (e.g.
+        ``"dispatch"`` or ``"combine"``), or ``None`` when nothing matches.
+    """
+    if not isinstance(candidate, dict):
+        return None
+    haystack = _candidate_haystack(candidate)
+    role_haystack = _role_haystack(candidate)
+    if not haystack or not role_haystack:
+        return None
+    for playbook in load_vendor_operator_playbooks():
+        match = playbook.get("match")
+        if not isinstance(match, dict):
+            continue
+        any_markers = [str(m).lower() for m in (match.get("any_marker") or [])]
+        if any_markers and not any(marker in haystack for marker in any_markers):
+            continue
+        name_markers = [str(m).lower() for m in (match.get("name_any") or [])]
+        matched_role = next((m for m in name_markers if m in role_haystack), None)
+        if name_markers and matched_role is None:
+            continue
+        result = copy.deepcopy(playbook)
+        result["role"] = matched_role or ""
+        return result
+    return None
+
+
+def playbook_group_id(playbook: dict[str, Any]) -> str:
+    """Return the stable group id a playbook's sibling roles share."""
+    return str(playbook.get("id") or "")
+
+
+def resolve_kernel_anchor_path(playbook: dict[str, Any]) -> str:
+    """Return a stand-in ``source_file`` path for a vendor-playbook candidate.
+
+    A vendor-playbook candidate has no rewritable device source, but
+    downstream tooling (``kernel_optimization.py``'s CLI, in particular)
+    still gates on a non-empty, path-shaped ``source_file`` before it will
+    dispatch to a backend at all. Point that field at the task bundle's
+    ``kernel_anchor`` file instead of leaving it empty -- resolved to an
+    absolute path under ``$FORGE_PATH`` when that's set (regardless of
+    whether the file exists on this host yet), else an absolute path under a
+    fixed, obviously-synthetic root so the value is still path-shaped (it
+    survives ``looks_like_source_path`` even when this analysis runs on a
+    host without the KernelForge checkout) without ever being a bare
+    relative string a later ``Path(...).resolve()`` could reinterpret
+    against an unrelated CWD.
+
+    Args:
+        playbook: A matched playbook entry (as returned by
+            ``match_vendor_operator_playbook``).
+
+    Returns:
+        An absolute path string; never empty as long as the playbook
+        declares a ``kernel_anchor``, and never relative -- a relative
+        string here would later be reinterpreted by ``Path(...).resolve()``
+        against whatever the apply-stage process's CWD happens to be, not
+        against this bundle (PR #1191 review finding #8).
+    """
+    anchor = str(playbook.get("kernel_anchor") or "").strip()
+    bundle = str(playbook.get("task_bundle") or "").strip()
+    if not anchor:
+        return ""
+    relative = f"{bundle}/{anchor}" if bundle else anchor
+    forge_root = (os.environ.get("FORGE_PATH") or "").strip()
+    if forge_root:
+        candidate = Path(forge_root) / relative
+        # Returned even when the file isn't there yet: this analysis host
+        # may lack the KernelForge checkout the apply stage will actually
+        # run against, but the path must still be absolute and anchored at
+        # a real, known root rather than silently falling through to a
+        # bare relative string.
+        return str(candidate)
+    # FORGE_PATH unset: still shape-check as path-like (needed to clear
+    # looks_like_source_path) without letting a bare relative string survive
+    # to be misresolved against an unrelated CWD later in the pipeline.
+    return str(Path("/nonexistent-forge-path") / relative)