tune: raise Build+OpenCode's step budget from 50 to 200 - #139

Summary

• Watched a real build (a from-scratch RLP codec, kimi-k2.7-code) run out of its 50-step budget one or two turns from done, via the persisted one-shot trail (fix(tui): persist one-shot session trails for post-mortem debugging #138): the implementation was already correct and independently verified, and the model was mid-fix on a syntax error in the test file — it had just run a python one-liner to inspect the exact broken byte range — when `[max_steps reached]` cut it off.
• 200, not unbounded: unlike the VM's own step ceiling (a compute-time guard, safe to raise arbitrarily high for trusted code — see `bin/lex-code`, fix(cli): raise the VM's step budget for lex-code's own entry points #137), each step here is a paid API call. An actually-stuck loop should still hit a wall rather than run forever.
• Scoped to `opencode_agent()` specifically, since that's what today's evidence covers — the other cloud providers share the same 50 by convention but haven't been observed hitting it.

Test plan

• `lex check` on the modified file, plus a full repo sweep
• `lex fmt --check src/`, `lex test` (4/4), `lex doc-sync --check`
• Will re-run the same RLP build task after merge to confirm the extra budget closes the gap

🤖 Generated with Claude Code