[fengshaoyi-amd]

What this does

Four functions were over the code-metrics gate at base, and none had a baseline entry, so the gate excused them as backlog: _resume_session (170 lines) and _seed_fresh_session (164) in cli/__init__.py, _restart_multinode (107) in multi_node/cli.py, and FieldDeclaration.__post_init__ (cognitive 32) in hyperloom_kb/schema.py. This PR brings each one under every limit, mostly by deleting what it did twice: status checks on a _short_poll result that is always SUCCEEDED, a second write of every restart key from a bag built for that one call, guards on values the function had already settled, and three type checks that FieldDeclaration.from_dict already makes. What is left of the two session functions splits along the jobs they ran in sequence.

Production code: +180 / −278 (net −98) across 3 files. Tests: none changed. The code-metrics baseline lowers one entry, module-lines of multi_node/cli.py (1,867 → 1,804), and gains none.

This continues the complexity cleanup of #2024, #2026, #2052, #2056 and #2057. The 6 commits are grouped by module and read one at a time; each commit message lists what it deleted and why that is safe.

Commit	Module	Production lines	Baseline file
8c58162 drop the terminal-failure checks _short_poll already makes	multi_node/cli.py	+8 / −27
d1e78d1 write each restart key once in _restart_multinode	multi_node/cli.py	+21 / −68
46dbf24 split _resume_session into opening the session and opening its next leg	cli/__init__.py	+72 / −96
4e17daa split _seed_fresh_session into resolving the target GPU and creating the session	cli/__init__.py	+79 / −81
c5f41f6 lower the multi_node/cli.py module-lines baseline to 1804	baseline only		1 lowered
72ab28d delete the type re-checks FieldDeclaration repeats from from_dict	hyperloom_kb/schema.py	+0 / −6

cli/__init__.py and multi_node/cli.py are under src/hyperloom/inference_optimizer/; schema.py is under src/hyperloom_kb/.

Gate numbers per function

Measured with the gate's own collectors (ruff C901, complexipy, function lines), origin/main → this branch. Limits are 20, 30 and 80; bold values were over the limit at base.

Module	Function	Cyclomatic	Cognitive	Lines
cli/__init__.py	_resume_session	14 → 3	19 → 5	170 → 48
	_seed_fresh_session	16 → 7	27 → 11	164 → 69
multi_node/cli.py	_restart_multinode	10 → 9	18 → 15	107 → 74
schema.py	FieldDeclaration.__post_init__	16 → 13	32 → 28	36 → 30

Also changed, never over at base (cyclomatic / cognitive / lines), all in multi_node/cli.py: _persist_multinode_restart 2 / 8 → 7 / 25 → 17; _resume_prior_launch 6 / 7 / 44 → 41; _start_pd_router 5 → 4 / 5 → 3 / 39 → 35; _submit_and_collect_pod_json 5 → 3 / 4 → 2 / 32 → 27; _restart_single_node 4 / 5 → 4 / 41; cmd_verify 3 / 2 → 1 / 36.

New functions, cyclomatic / cognitive / lines, all in cli/__init__.py: _open_resumed_session 7 / 6 / 50, _open_next_leg 6 / 7 / 44, _resolve_target_gpu 6 / 13 / 34, _create_fresh_session 2 / 1 / 55. Each takes one to three parameters.

Deleted: the _MultinodeLaunch dataclass in multi_node/cli.py.

Across the 3 touched production files: C901 total 743 → 737, cognitive total 1,025 → 1,008, functions 196 → 200. Functions over a gate limit: cognitive 2 → 1, length 9 → 6. What is left is outside this work: in cli/__init__.py, _run_optimize (279 lines, cognitive 43), _reexport_session_environment (111), _export_partition_shape (101), _resolve_run_topology (92) and _resolve_agent_lineup (89); in multi_node/cli.py, build_parser (266). Open PRs change _run_optimize (#1450, #1546, #2062), _reexport_session_environment (#1979, #2055) and _resolve_run_topology (#1924), so this PR leaves them alone.

Module lengths: cli/__init__.py 2,778 → 2,752, multi_node/cli.py 1,870 → 1,804, schema.py 1,014 → 1,008.

What changes in each module

Multi-node restart (multi_node/cli.py)

Removed:

• The status re-checks that five _short_poll callers made on the job it returned: a fail-fast block in _restart_multinode, a "router job ended" exit in _start_pd_router, two EXIT_TRANSIENT returns in _submit_and_collect_pod_json, and a log print in _restart_single_node and in cmd_verify.
• The _MultinodeLaunch dataclass and the second write of the restart keys it carried. _persist_multinode_restart(state, args, *, kill_sub, router_sub, router_state) now writes only what the checkpoint leaves until the poll returns: the kill submission and, for a disaggregated launch, the PD fields, the router submission and the leg URLs. It takes five parameters instead of three; the bag's other four fields (the PID and log directories, the launch submission and the topology record) were keys the checkpoint had already written.
• The re-derivation of tp, ep, pd_mode and extra_args from args at each use. The checkpoint writes them from the topology record.
• The prev_sub parameter of _resume_prior_launch, which now reads the prior submission id from the state it already takes.

Session start (cli/__init__.py)

New names:

• _open_resumed_session: check that --resume-from names a session under the workspace root, pin it, take its lock, run the resume guard, and load its manifest once state.json is known to exist.
• _open_next_leg: refuse a target_reached stop without --force-resume, then begin the leg, grant --extend-hours, record the model-gate skip and print the restored budget.
• _resolve_target_gpu: settle args.gpu_type from the probe and the --gpu-type hint, export the Magpie runner, refuse a framework that runner cannot serve, and check that provenance can name the ISA.
• _create_fresh_session: create and lock the session directory, then write its manifest, launch info, partition shape, GPU power settings and seeded state.

_resume_session now runs the first two in order, and _seed_fresh_session the last two.

Removed:

• Resume: the __import__ hop to reach _SESSION_SKELETON; the hand-written containment check, now session.paths.is_path_within, which owns that rule; a one-element stop-reason set and a --force-resume re-test that could not fail; two (state.current_best or {}); and four hand-written "state value, else manifest value" copies, now one comprehension over the same keys in the same order.
• Fresh session: a second args.gpu_type = None and a runner_gpu_type or gpu_type fallback for a board without a Magpie runner; the empty-model fallbacks in the launch info and the partition model; and three copies of "exit 2 if this preflight fails", now one condition whose or keeps the short circuit.
• Comments that restated a docstring or narrated the next line, and a "no backfill needed" tombstone. Every comment that states a constraint moved with its code or into the docstring of the function that now owns it. The Langfuse marker's except Exception moved into _create_fresh_session unchanged; no new handler was added.

KB field declaration (hyperloom_kb/schema.py)

FieldDeclaration.__post_init__ no longer re-validates the three attributes FieldDeclaration.from_dict already converts or rejects before it constructs: kind as a FieldKind, role as a FieldRole, and required, sensitive, many and group as booleans. Every check on a value that from_dict passes through raw stays: the name, the description, the declared values, every combination rule, and the search weight. The weight is still normalized to a float, so a declaration with an integer weight keeps one schema_ref across a round trip.

Kept at base: sweep_case

sweep_case (src/kernelforge/evaluation/benchmark/runner.py:581) is over the parameter limit with seven keyword parameters and stays at seven; its function-args baseline entry is unchanged. Its callers:

• Production, src/:
  • runner.py:707-716, the python3 -m kernelforge.evaluation.benchmark.runner CLI (_sweep_cli). It passes all seven, taking warmup_iters, bench_iters and timeout_sec from --warmup, --iters and --timeout (:694-696).
  • kernelforge/orchestrator/specialist_probe.py:394-400 and :450 (probe_variant, then _measure_on_device). They pass driver_script, case_id, constants, prefix_constants and timeout_sec. The probe resolves sweep_case by name (:41-43, :197-200), and it refuses a primitive that does not accept those five keywords (:45-51, :218-231).
• Production, scripts/: none.
• Tests only: kernelforge/tests/evaluation/benchmark/test_sweep_case.py:42-44, whose call sites pass case_id, constants, timeout_sec and prefix_constants; and the probe stand-ins in kernelforge/tests/test_specialist_probe.py:401, :414 and :740, which have the same five-keyword signature.

No parameter can be derived from another, and warmup_iters and bench_iters are the only two that the probe and the tests never pass. The only way to five parameters without bundling arguments therefore deletes the --warmup and --iters CLI flags, a CLI interface change this PR does not make. Passing the existing BenchRun in place of driver_script, warmup_iters and bench_iters would only bundle arguments, and the probe's keyword check requires driver_script by name.

Why the deletions are safe

Line numbers are on this branch.

• Status re-checks after _short_poll. _short_poll (multi_node/cli.py:267-300) returns a job only when its status is in _TERMINAL_OK_STATUSES, which is {"SUCCEEDED"} (:60, returned at :290-291). FAILED and STOPPED raise RuntimeError (:292-293), and a timeout raises TransientFailure (:295-299), so no caller can see another status. A FAILED or STOPPED job's RuntimeError reaches the multi-node CLI's dispatcher handler, except (RuntimeError, ValueError) in main (multi_node/cli.py:1786); in-process, the orchestrator catches it around cmd_restart_server and cmd_apply_tracelens_patch (orchestrator/actions/executors/_multi_node_server_lifecycle.py:507, :400). So the job logs those deleted branches meant to print on a failure were never printed, before or after this PR. Printing them would be a behaviour change, which this refactor does not make.

• Restart keys. The early checkpoint (multi_node/cli.py:1302-1318) writes the PID and log directories, the launch submission, the framework, the model, tp, ep, pd_mode, extra_args and the topology record before the poll. tp, ep, pd_mode and extra_args come from the record _rayjob_topology_fingerprint builds (:1120-1128), with the expressions the checkpoint used at base. _persist_multinode_restart (:1258-1274) runs after the poll and the PD router, and writes only the keys the checkpoint cannot have yet. _resume_prior_launch (:1174-1214) reads the prior submission id from state (:1187), the value the deleted parameter carried.

• Resume. is_path_within (session/paths.py:220-226) is the containment rule the hand-written check copied. SharedState.from_dict always yields a dict for current_best (orchestrator/state/shared_state.py:1103-1104).

• Fresh session. Only an empty GPU type reaches the branch for a board without a Magpie runner, and args.gpu_type is already None there. The function has exited on an empty model before it writes the launch info and the partition model, and the quantization prelude only replaces args.model with a Path (cli/quantization.py:80). The preflight condition (cli/__init__.py:2197-2202) runs the three checks in the order the three copies did and stops at the first that fails.

• KB field declaration. FieldDeclaration.from_dict (schema.py:468-495) converts or rejects kind (:476), role (:481) and the four flags (:488-491) before it constructs. Every declaration that comes from outside Python code reaches the dataclass only through ExperienceDeclaration.from_dict (schema.py:700) and then FieldDeclaration.from_dict, which is the path for:

  • declaration YAML and JSON files (hyperloom_kb/config.py:35, also behind collect/mapping.py:162);
  • KB database rows (http_service.py:449);
  • HTTP PUT bodies (http_service.py:1262);
  • spool files (remote.py:506);
  • remote export responses (remote.py:567).

  In the repo, the only direct constructors are RATIONALE_DEFAULTS (schema.py:500-502) and tests, and they pass FieldKind and FieldRole members and bool literals.

Observable effect

Nothing an operator sees changes. For every input the repo's own callers produce, these are the same, in the same order:

• stdout and stderr lines and exit codes;
• state.json, the session manifest and the launch info;
• the multi-node state file and the exported environment;
• stored KB declarations, their to_dict output and their schema_ref. An invalid kind, role or flag in a declaration document still raises the same SchemaValidationError from from_dict, before anything is constructed.

Fake-driven differential runs compared base and head with no difference:

• _restart_multinode: 12 scenarios (aggregated and PD, fresh, resumed RUNNING, resumed serving, router without leg URLs, poll timeout, --print-logs). Saved state snapshots, log lines and return codes match.
• _resume_session: 13 scenarios (plain, target_reached with and without --force-resume, other stop reasons, --extend-hours, outside the workspace, missing directory, resume blocked, no manifest, no state.json, stale AgentX state, resume conflict, null current_best). Output, exit codes, call order and state.json match.
• _seed_fresh_session: 20 scenarios (missing model, model from MODEL_PATH, local and remote probe, a probe that disagrees with the hint, no GPU, atom, framework from FRAMEWORK, quantized model, AgentX, --profile-osl, no framework version, each preflight failing, Langfuse failing, a family-runner board, R9700 with vLLM). Output, exit codes, call order, exports and args match.
• FieldDeclaration: 1,456 cases, every combination of kind, role, flags, values and search weight built in Python, plus valid and invalid from_dict documents. Results and error messages match.

The differences below are reachable only by input that no caller in the repo produces:

• restart-server with a tp that is not an int.
  • Before: last_restart_tp in the multi-node state ended as the raw value, because the persist step overwrote the checkpoint's int(args.tp).
  • Now: it stays int(args.tp).
  • Callers: the CLI parses --tp with type=int (multi_node/cli.py:1536). The in-process caller passes tp=tp_int (orchestrator/actions/executors/_multi_node_server_lifecycle.py:446). Tests pass tp=8.
• A _short_poll that returns a job whose status is not SUCCEEDED. Only a test double can; the real one raises instead (multi_node/cli.py:288-299).
  • Before:
    • for FAILED or STOPPED, _restart_multinode and _start_pd_router printed the job logs, logged an ERROR line and failed with exit code 1, and cmd_verify and _restart_single_node printed the job logs;
    • for any status but SUCCEEDED, _submit_and_collect_pod_json returned EXIT_TRANSIENT.
  • Now: each continues as for a SUCCEEDED job, and prints logs only with --print-logs.
  • Callers: the only test that replaces _short_poll returns SUCCEEDED (inference_optimizer/tests/test_external_multi_node.py:941).
• --resume-from under a workspace root whose resolve() raises OSError or RuntimeError.
  • Before: a traceback from the containment check.
  • Now: is_path_within answers no, so the operator gets the existing "is not under $USER_DATA_PATH" error and exit 2.
  • Callers: the root comes from the operator's $USER_DATA_PATH through workspace_root() (session/paths.py:85). No caller or test in the repo sets one that cannot be resolved.
• A resolved GPU type that is not empty but is blank.
  • Before: args.gpu_type became None, and the custom-entrypoint check received the blank string.
  • Now: the blank string stays in args.gpu_type, and the check receives "".
  • Callers: the type is probed or user_specified (gpu_types.py:122). Each comes from a resolver that returns a known board tag or None: _resolve_amd_gpu_type (gpu_types.py:153-162), _autodetect_gpu_type (gpu_types.py:125-150) and remote_autodetect_gpu_type (multi_node/_internal/gpu_probe.py:56-70, through _parse_gpu_type at :41-53).
• Python code that constructs FieldDeclaration directly with a value its annotations forbid (the package ships py.typed):
  • A kind that is not a FieldKind, for example the string "number". Before: SchemaValidationError("field 'x' has an invalid kind"). Now: a later combination rule may raise its own SchemaValidationError, for example "only a string field declares values" for "string" with values; otherwise the object is built, and to_dict() and schema_ref raise AttributeError.
  • A flag that is not a bool, for example required="yes" or many=1. Before: SchemaValidationError("field 'x' required must be a boolean"). Now: the value is accepted and read for its truthiness, and to_dict() emits it as given. from_dict rejects it, so the declaration cannot survive a round trip.
  • A role that is not a FieldRole, for example the string "decision". Before: SchemaValidationError("field 'x' has an invalid role"). Now: with many=True, "a field with a role holds one value" is raised; otherwise the object is built, and to_dict() and schema_ref raise AttributeError.

Checklist

• Description: above.
• Linked issue(s): none.
• Tests:
  • No test was added or changed.
  • The affected tests ran on Windows at 958874f, base against head. origin/main has since gained refactor(breakdown, kernelforge): delete what eleven over-gate functions re-checked #2052, a refactor of breakdown/ and kernelforge/rewrite_by_flydsl/ that changes none of the three files or the tests below.
  • Session start and multi-node restart: the 26 test files that reach the three functions, with -n 2, give 1,041 passed and 11 skipped on both sides, and the same 121 failed and 76 errors, with no failing id on one side only. These fail on base too, on Windows only, mostly in test_apply_gpu_power_settings.py, test_multi_node_scripts.py and test_external_multi_node.py.
  • On Windows the multi-node restart tests never reach _restart_multinode, because os.getuid is missing in state_file_safe_to_read. A local shim for that one call, used the same way on both sides, lets five multi-node files run it: 266 passed and the same 15 failures (POSIX group checks) on base and head.
  • KB declarations: the 24 test files that construct or load declarations (the 21 files of hyperloom_kb/tests plus test_experience_collect.py, test_experience_kb_integration.py and test_experience_kb_service.py) give 280 passed, 1 skipped and 19 failed on head. The same 19 fail on base: local-service startup (an SSL error on this host), SIGTERM, directory fsync and file-path tests.
  • Diff coverage: cli/__init__.py and multi_node/cli.py are on the coverage omit list (pyproject.toml, [tool.coverage.run] omit, :360-361), so diff coverage has no countable statements for them, and the schema.py change only deletes lines.
  • Linux CI on this PR is the authority.
• Size/complexity triggers crossed:
  • 458 changed production lines (+180 / −278) is over the 400-line warning and under the 1,000-line budget, so this does not carry size-exception.
  • No added or changed function crosses a gate limit. _persist_multinode_restart takes five parameters, at the limit (three before, one of them the _MultinodeLaunch bag); the new functions take one to three.
  • Two functions are over the ~60-line trigger and under the 80-line gate. _restart_multinode (74, from 107) is the one restart sequence, in its order: kill or resume, launch, the early checkpoint, the poll, the PD router and the persist. _seed_fresh_session (69, from 164) is the fresh-session steps in their fixed order (model, quantization, framework, target GPU, workload exports, session creation, model gate and preflights), each a call or the environment exports it publishes for subprocesses.
  • One function is over the cyclomatic-10 trigger: FieldDeclaration.__post_init__ (13, from 16), one if per declaration rule that from_dict does not already enforce.
  • cli/__init__.py (2,752) and multi_node/cli.py (1,804) stay over the 1,200-line module limit they were already over; each gets shorter, and cli/__init__.py stays over its 2,569 baseline entry, as at base. schema.py (1,008) is on the report-only 800-line warning list.
• If this simplifies or refactors:
  • Mechanisms removed: listed per module under What changes in each module.
  • Contract preserved: the CLI's output, exit codes and session files, the multi-node state file, every in-repo caller, and the stored KB declarations and the errors from_dict raises.
  • Tests: as above.
• Observable effect: none; see the section above for why, and for the out-of-contract inputs that now behave differently.
• Breaking changes: no. The changed signatures (_persist_multinode_restart, _resume_prior_launch) belong to private functions whose only callers are in the same file. FieldDeclaration keeps its fields, and from_dict keeps its errors.
• PR addresses single concern: yes, bringing four over-gate functions under the gate by deleting what they did twice. The KB declaration change rides along because it is six deleted lines of the same kind, checks repeated from the one place that already makes them, too small to take its own CI round and review.
• Root cause is upstream: no.

Known merge interactions

Each open PR that changes one of these files was trial-merged with git merge-tree --write-tree, once into origin/main alone and once into this branch. A conflict that also occurs against main alone is not introduced by this branch.

• Open PRs that change cli/__init__.py: feat(trace): per-LLM-call attribution, and the reports that read it #1450, feat(explore): let a scriptable adapter declare its own lever grid #1546, xDiT profiling and TraceLens analysis support #1762, Migrate from the deprecated rocm-smi to amd-smi #1924, feat(specialist): add a speculative decoding specialist type #1979, [fix] close the defects behind the MuseSpark run shortfall and delete state nothing reads #2055 and Run the orchestration role on the DeepSeek Harness #2062. None gains a conflict from this branch, in this file or anywhere else.
  • feat(trace): per-LLM-call attribution, and the reports that read it #1450 and feat(explore): let a scriptable adapter declare its own lever grid #1546 already conflict once in cli/__init__.py against main, and the block is identical against this branch. feat(trace): per-LLM-call attribution, and the reports that read it #1450 conflicts in 23 files and feat(explore): let a scriptable adapter declare its own lever grid #1546 in 2, either way.
  • xDiT profiling and TraceLens analysis support #1762 (draft) merges cli/__init__.py cleanly. Its 4 conflict blocks in the baseline file occur either way; in one of them, this branch's side reads multi_node/cli.py 1804 where main's reads 1867.
  • Migrate from the deprecated rocm-smi to amd-smi #1924 (draft) conflicts in 28 files either way, none of them this PR's.
  • feat(specialist): add a speculative decoding specialist type #1979 and [fix] close the defects behind the MuseSpark run shortfall and delete state nothing reads #2055 merge cleanly.
  • Run the orchestration role on the DeepSeek Harness #2062 has one conflict in the baseline file, identical against main.
• Add RDNA4 GPU support (RX 9070 XT, RX 9070, RX 9060 XT, RX 9000 XT) f… #1032, Add xDiT kernel source mapping support #1154 and Set --pids-limit in docker-mode launch templates (fixes silent hangs + OpenBLAS crashes under Ray + TP>1) #1276 are stale and change both inference-optimizer files. They conflict in 779, 891 and 958 files either way. In cli/__init__.py their 25, 49 and 63 blocks are identical either way. In multi_node/cli.py they have 22, 29 and 29 blocks either way, three of which this branch's restart changes reshape in place.
• No open PR changes hyperloom_kb/schema.py.
• 49 other open PRs share only scripts/code_metrics_baseline.txt. 14 of them conflict there against main alone, with the same blocks against this branch; in fix: judge explore variants and warm recipes by what they change about the running launch #1908, one block's side reads 1804 here instead of 1867. One conflicts only against this branch: refactor(inference-optimizer): delete the dead and repeated code in ten over-gate CLI, roofline and multi-node functions #2077 (cc20, lean/shafeng/cc20-io-misc-pr), whose lowered module-lines rows for model_gate.py, preflight.py and roofline_ceiling.py sit next to this PR's multi_node/cli.py row. Keeping both sides' rows resolves it.

Gate

python scripts/code_metrics.py --base-ref origin/main (origin/main at e25cb2d) passes on the 4 touched files: 0 new, 0 worse and 0 out-of-date entries in every dimension, no new or worse duplication or dead code, and 0 findings from the comment, English-only, test-import, repeated-literal and A1–A12 checks. Baseline: one module-length entry lowered (multi_node/cli.py 1,867 → 1,804), +0.

The gate lists under Already over at the base: _run_optimize (279 lines), _reexport_session_environment (111), _resolve_run_topology (92) and _resolve_agent_lineup (89), and cli/__init__.py module length 2,778 → 2,752 (baseline entry 2,569). schema.py (1,008 lines) is on the report-only module-length warning list. ruff check and ruff format --check are clean on the three files.

[github-actions]

Code metrics gate: PASSED

Judged: the 4 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.

Dimension	Limit	New	Worse	Baseline out of date	Baseline entries	vs base
Cyclomatic complexity	fails > 20	0	0	0	55	+0
Cognitive complexity	fails > 30	0	0	0	203	+0
Function length (lines)	fails > 80	0	0	0	346	+0
Nested block depth	fails > 5	0	0	0	20	+0
Parameter count	fails > 5	0	0	0	95	+0
Module length (lines)	warns > 800, fails > 1200	0	0	0	49	+0
Duplicated lines per file	fails > 0	0	0	0	96	+0
Dead code (vulture)	fails > 0	0	0	0	2	+0

Check	Findings	Effect
Comments	0
English only	0
Production code imports test code	0
Repeated literals	0
A1 Private name of another package	0
A8 Flag argument	0
A9 Tuple return	0
A10 Raw vocabulary comparison	0
A11 Hidden global write	0
A12 Pure forwarder	0
A12 Single-implementation abstraction	0

Module length warning: over 800 lines, within the 1200-line failure limit (not failing): 1

Module	Lines	Limit
src/hyperloom_kb/schema.py	1008	warns > 800

Already over at the base (backlog, not this change's): 5

Unit	Dimension	Base	Now
src/hyperloom/inference_optimizer/cli/__init__.py:1782 _reexport_session_environment	Function length (lines)	111	111
src/hyperloom/inference_optimizer/cli/__init__.py:2312 _resolve_agent_lineup	Function length (lines)	89	89
src/hyperloom/inference_optimizer/cli/__init__.py:2207 _resolve_run_topology	Function length (lines)	92	92
src/hyperloom/inference_optimizer/cli/__init__.py:2419 _run_optimize	Function length (lines)	279	279
src/hyperloom/inference_optimizer/cli/__init__.py:1	Module length (lines)	2778	2752

98 findings in files this change does not touch

Outside this change (informational): 98

Unit	Dimension	Baseline	Now
src/hyperloom/orchestrator/state/shared_state.py resolve_graded_comparison	Cognitive complexity	absent	34
src/hyperloom_kb/collect/mapping.py _experience	Cognitive complexity	absent	31
src/kernelforge/cli/forge_loop_command.py _prepare_measurement_driver	Cognitive complexity	absent	36
src/kernelforge/cli/forge_loop_command.py _read_kb_warm_start	Cognitive complexity	absent	40
src/kernelforge/cli/forge_loop_command.py forge_loop	Cognitive complexity	absent	31
src/kernelforge/loop/runner.py IterationLoop._collect_session_candidate	Cognitive complexity	absent	44
src/kernelforge/loop/runner.py IterationLoop._rule_on_attempt	Cognitive complexity	absent	38
src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py _build_parser	Function length (lines)	absent	84
src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py _spawn_rank_actors	Function length (lines)	absent	84
src/hyperloom_kb/sync.py GlobalSync.pull	Function length (lines)	absent	84
src/kernelforge/cli/forge_loop_command.py _RunReporter.build_result	Function length (lines)	absent	92
src/kernelforge/cli/forge_loop_command.py _prepare_measurement_driver	Function length (lines)	absent	136
src/kernelforge/cli/forge_loop_command.py _read_kb_warm_start	Function length (lines)	absent	149
src/kernelforge/loop/runner.py IterationLoop._close_iteration	Function length (lines)	absent	117
src/kernelforge/loop/runner.py IterationLoop._collect_session_candidate	Function length (lines)	absent	185
src/kernelforge/loop/runner.py IterationLoop._implementer_history	Function length (lines)	absent	112
src/kernelforge/loop/runner.py IterationLoop._rule_on_attempt	Function length (lines)	absent	102
src/kernelforge/loop/runner.py IterationLoop._source_attempt	Function length (lines)	absent	175
src/hyperloom_kb/http_service.py	Module length (lines)	absent	1525
src/kernelforge/cli/forge_loop_command.py	Module length (lines)	absent	2061
src/hyperloom/orchestrator/actions/executors/_geak_sweep.py sweep_via_geak	Cognitive complexity	67	70
src/hyperloom/orchestrator/actions/executors/explore.py ExploreExecutor.__call__	Cognitive complexity	369	383
src/hyperloom/orchestrator/kernel/request_handlers.py _run_forge_gemm_tuning	Cognitive complexity	124	125
src/hyperloom/orchestrator/phases/kernel.py KernelPhase._run_geak_kernel_phase	Cognitive complexity	122	130
src/hyperloom/orchestrator/phases/prelude.py PreludePhase.maybe_enqueue_warm_replay	Cognitive complexity	190	191

...and 73 more; the job log lists every row.

Verdict computed by the gate scripts in /home/runner/work/_temp/code-metrics-gate, not by this tree's copy.

_{Thresholds and sources: [tool.hyperloom.code_metrics] in pyproject.toml. Tools: complexipy 8.0.1, jscpd 5.4.1, ruff 0.16.2, vulture 2.16.}

[ZhengGong-amd]

Reviewed: gate numbers re-measured (ruff C901, complexipy, function lines) and match the description; each removed guard / "producer always sets it" claim checked against its producers and call sites; base vs head differential runs show no undisclosed behaviour change; tests covering the touched modules pass.

LGTM.