Migrate hand-crafted agents (Gemmini, CUDA, trn-nki1/2, Saturn) to Bu…

[Bryce-Lim]

Refs #31

What this does

Hand-ports the prompts from each LLMAgent subclass in autocomp/agents/
into .md / .yaml configs under autocomp/agent_builder/.built/, so they
can run through BuiltLLMAgent instead of needing custom subclasses.

This is step 1 of the linked issue. The legacy LLMAgent subclasses are
left wired up so existing experiments still work; step 2 (running the Agent
Builder for each backend and comparing against these hand-migrated prompts)
is a separate follow-up.

New built-agent directories

• built:gemmini (gemm/conv, pe_dim != 4)
• built:gemmini-admm (admm-multifunction, pe_dim == 4)
• built:cuda
• built:trn-nki1
• built:trn-nki2
• built:saturn

How to verify

for a in gemmini gemmini-admm cuda trn-nki1 trn-nki2 saturn; do
  python -m autocomp.agent_builder.run_agent_builder \
    --agent-name $a --inspect autocomp/agent_builder/.built/$a
done

Then a 1-iteration run with agent_name = "built:gemmini" vs.
agent_name = "gemmini" to spot-check that the prompts look comparable.

Things worth a closer look during review

• trn-nki1/trn-nki2 architecture.md were partly drafted by an LLM
  helper; I caught and fixed wrong SBUF/PSUM sizes but the rest of those
  paragraphs deserve a once-over.
• trn-nki2/isa_docs.md has fewer ### subsections (25) than
  trn-nki1/isa_docs.md (91). Could be a real difference in the v2
  generator's coverage or could be an over-aggressive trim — worth
  diffing against autocomp/agents/trn_nki2/nki_isa_generator.py.
• Saturn's code_examples.md is intentionally a stub (the legacy Saturn
  agent didn't include ICL examples).
• One small behavior fix in built:saturn/rules.yaml: the legacy Saturn
  rule still says Optimize the test() function, but the project renamed
  that entry point to solution() on 4/6/2026. The migrated copy says
  solution(). Revert if the legacy test() was intentional.

[Bryce-Lim]

@charleshong3 did some work in the past week surrounding issue #31 , would be greatly appreciated if you could take a look!

[charleshong3]

I think the Trainium and Saturn agents aren't needed due to them already having built agents. They could be useful for the research so maybe keep them on your fork, but I would remove them from this PR so we don't clutter the repo.

For the CUDA agent, maybe make the architecture and ISA docs specific to a particular GPU, for example the L40S (which we have in the lab)? Information that is generic across GPUs is probably already well-represented in the LLM pretraining data.