Fix per-domain eval args misalignment when polyglot is combined with other domains

[rootkiller6788]

Problem

generate_loop() evaluates polyglot through a separate harness (run_harness_polyglot), so it passes a polyglot-filtered domain list into generate():

generate(docker_client, [d for d in domains if d != "polyglot"], ..., eval_samples=eval_samples, eval_subsets=eval_subsets, ...)

but it passes the full eval_subsets / eval_samples lists unchanged. Inside generate(), the staged and full evaluation loops pair them positionally:

zip(domains, eval_subsets, stagedeval_samples)

Whenever polyglot is not the last domain in --domains, the domain list is shorter than the argument lists, so every domain after polyglot is paired with polyglot's eval subset and sample count, and the remaining domains shift by one.

Reproduction

python generate_loop.py --domains polyglot search_arena genesis_go2walking --eval_samples 10 20 30

Before this change, the generate() call for that run receives:

domains=['search_arena', 'genesis_go2walking']
eval_samples=[10, 20, 30]      # search_arena gets 10 (polyglot's), genesis gets 20 (search_arena's)
eval_subsets=['', '_filtered_100_train', '']   # search_arena gets polyglot's empty subset

So search_arena would be evaluated on the wrong subset/sample count, and genesis_go2walking would inherit search_arena's arguments. When polyglot is last (or absent), zip() truncation masks the problem, which is why it only shows up with polyglot in a non-final position.

Fix

Filter the per-domain evaluation arguments alongside the domain list so each remaining domain keeps its own eval_subset / eval_samples:

non_polyglot_domains = [d for d in domains if d != "polyglot"]
non_polyglot_eval_subsets = [s for d, s in zip(domains, eval_subsets) if d != "polyglot"]
non_polyglot_eval_samples = [n for d, n in zip(domains, eval_samples) if d != "polyglot"]

and pass those to generate(). Runs without polyglot are unaffected (the lists pass through unchanged).

Verification

• A mocked integration harness that runs generate_loop() (heavy dependencies stubbed) confirms the pre-patch generate() call is misaligned and the post-patch call receives correctly aligned per-domain arguments.
• python -m py_compile generate_loop.py passes.