exp: terminal-bench-2-v1 baseline results in README

[mikasenghaas]

Summary

Adds examples/tasksets/terminal_bench_2_v1/README.md (the taskset had none) with baseline results for three models on terminal-bench-2-v1 — same command per model, rlm harness on the modal runtime, served via Prime Intellect inference, n=1 rollout/task (89 tasks):

uv run eval terminal-bench-2-v1 --harness.runtime.type modal --harness.id rlm \
  --max-turns 100 --timeout.rollout 3600 -m <model>

Model	Accuracy	Corrected (excl. caps)	Reward	Errored	Capped	MC-400s	Wall
deepseek/deepseek-v4-flash	48.3% (43/89)	59.7% (43/72)	0.483	0	17	23	~60 min
z-ai/glm-4.7	29.2% (26/89)	33.8% (26/77)	0.292	0	12	72	~60 min
qwen/qwen3.6-27b	18.0% (16/89)	18.0% (16/89)	0.180	1	0	72	~48 min

• Corrected = accuracy excluding rollouts that hit the turn/time cap (none of the capped rollouts solved in any run).
• Capped = hit the 100-turn or 3600 s limit rather than the agent stopping itself — qwen never capped (always agent_completed), it just finished early and wrong.
• MC-400s = transient context-length 400s the rollout retried/recovered from (not terminal errors).

Note

Add baseline results README for terminal-bench-2-v1 taskset

Adds README.md documenting the terminal-bench-2-v1 taskset, including prerequisites (harbor CLI and container runtime), a command example for running with the rlm harness on Modal, and a results table with model accuracy and related metrics with explanatory footnotes.

^{Macroscope summarized 0660d4d.}