Add GEAK v4 kernel workflow agent integration

[irvineoy]

Summary

• add a registered geak_v4 agent that runs GEAK's deterministic kernel_workflow.js through a persistent Claude Agent SDK client
• add Docker setup, read-only GEAK checkout mounting, persistent SDK dependencies, preflight checks, MI300 quickstart config, and user documentation
• run GEAK on a disposable task snapshot and import only a Director-accepted patch for the one declared source file
• handle background Workflow lifecycle, disk artifact recovery, physical-to-logical GPU mapping, bounded output, and process-group cleanup

V1 task contract

The initial adapter supports hip2hip, triton2triton, and flydsl2flydsl tasks with exactly one ordinary source file. It rejects protected test/harness/config paths, symlinks, hard links, multiple sources, binary patches, path traversal, and create/delete/rename/mode or multi-file changes.

The adapter was developed against GEAK commit 4965d5b2ccde927925c8c5501a25c1233daa52eb (v4.0.0-102-g4965d5b).

Integrity boundary

• GEAK receives a disposable workspace copy and apply_to_original=false
• the scored workspace is fully manifested before and after Workflow execution
• Director, Workflow, and imported patch provenance must all match the pinned eval/final_patch.diff
• result/audit JSON uses symlink-safe atomic writes, artifact directories are inode-pinned, and normal runs mount the persistent Python dependency target read-only
• this is a fail-closed integrity boundary, not an OS sandbox; that limitation is called out in the agent README

Validation

• python -m pytest -q tests — 84 passed
• bash tests/test_docker_benchmark.sh — passed
• Python compile checks, shell syntax checks, YAML parsing, git diff --check, and Make target dry runs — passed
• handoff mapping against the real GEAK checkout and pinned kernel_workflow.js — passed
• cached gfx950 container setup/import smoke for claude-agent-sdk==0.2.127, including ClaudeSDKClient, ClaudeAgentOptions, and cli_path — passed without contacting Claude
• two independent read-only reviews found no remaining blocker after patch-provenance, lifecycle, dependency-mount, outer-Git-worktree, and symlink-write fixes

Not tested

A live Claude/GEAK optimization run was intentionally not executed because this machine has no paid Claude Code account. The included one-task quickstart is the remaining authenticated end-to-end check.

Suggested review focus

• compatibility of the handoff and normalized result schema with the pinned GEAK revision
• Claude SDK background-task lifecycle and terminal artifact semantics
• whether GEAK maintainers want to widen the V1 single-source task contract in a follow-up