coli: honor THINK=1 in run mode

[anrasi]

Summary

coli run hardcodes the nothink template — <|assistant|><think></think> — so THINK=1 silently has no effect in one-shot runs: the reasoning block is pre-closed before the engine sees a token. Serve mode already picks the template from the env in glm.c (getenv("THINK") && atoi(...) ? "<think>" : "<think></think>"). This PR makes cmd_run pick it the same way: two lines, default unchanged (nothink), and ENVIRONMENT.md's "THINK=1 turns on visible reasoning" now holds for run the way it does for serve.

Noticed on a DGX Spark GB10 while batch-running long planning prompts through coli run: outputs consistently opened with an empty <think></think> despite THINK=1 in the environment. With the patch the model emits its reasoning block; we've been running exactly this change locally for multi-hour THINK batches on the 744B model today.

coli chat needs no change — it hands the raw text to the engine, whose template already honors THINK.

Validation

• make -C c check — clean build, C unit tests, and the Python suite (73 tests) green on this branch (aarch64, gcc 13.3)
• CUDA — n/a (launcher-only change)
• No performance claims

Compatibility

• The default CPU build remains dependency-free (stdlib os.environ only)
• No model files, generated binaries, or benchmark artifacts

🤖 Generated with Claude Code