fix: remove worktree from PYTHONPATH to prevent system package shadowing

[iraj465]

During optimization, agents may create directories in their worktree that match system package names (e.g. aiter/). Because GEAK_WORK_DIR was placed first in PYTHONPATH, these shallow stubs shadowed the real system packages, causing ImportError during full_benchmark evaluation.

Remove GEAK_WORK_DIR from PYTHONPATH in both the COMMANDMENT SETUP template (run.sh / run_harness.sh) and build_eval_env(). The worktree is not needed on PYTHONPATH: harnesses load kernel.py via importlib with an explicit file path, and Python automatically adds the script directory to sys.path[0].

Example:

1. Preprocessing — GEAK sets up the kernel workspace:

worktree: /path/to/worktrees/slot_0/
├── kernel.py (the triton kernel to optimize)
├── test_kernel_harness.py
└── (nothing else)
2. Agent optimizes — the AI agent edits kernel.py, runs save_and_test. During optimization, the agent sometimes creates helper files. For fused_append_shared_experts, the kernel imports from sglang which imports from aiter. The agent sees import errors, tries to "fix" them by creating a local aiter/ package stub:

worktree: /path/to/worktrees/slot_0/
├── kernel.py
├── test_kernel_harness.py
└── aiter/ ← agent created this
├── init.py ← stub: just "from pkgutil import extend_path"
└── ops/
3. Round ends, full_benchmark runs — GEAK's evaluation.py runs the harness in a subprocess with:

export PYTHONPATH=/path/to/worktrees/slot_0:/repo_root:$PYTHONPATH
python3 test_kernel_harness.py --full-benchmark
4. Python resolves import aiter — it searches PYTHONPATH left to right:

/path/to/worktrees/slot_0/aiter/init.py ← FOUND FIRST (the stub!)
/sgl-workspace/aiter/aiter/init.py ← never reached
5. Import fails:

from aiter import dynamic_per_tensor_quant # stub doesn't have this → ImportError
Would we need this fix for any new kernel URL?
Yes, but only if the agent creates files that shadow a system package. The issue is general:

GEAK puts the worktree first in PYTHONPATH
If the agent creates any directory that matches a system package name (e.g., torch/, numpy/, triton/, aiter/), it will shadow the real package
This isn't aiter-specific — it could happen with any package
In practice, it happens most with aiter because:

Many kernels import from sglang → sglang imports aiter
The agent sees aiter import errors and tries to create a local package to "fix" them
But any kernel where the agent creates a directory matching a system package would hit this
The fix is general-purpose: removing the worktree from PYTHONPATH (or putting it last) prevents ANY package shadowing, not just aiter. The harness loads kernel.py via importlib with an explicit file path — it doesn't need PYTHONPATH to find it. So removing the worktree from PYTHONPATH has no downside

[sdubagun-amd]

@iraj465 Can you update the PR after addressing the merge conflicts and CI failures?