fix(core): clamp compaction output budget

[opencode-agent]

Summary

• clamp the output-token reservation used by V2 automatic compaction to 32k, matching V1
• prevent models advertising output equal to context (including Kimi K2.7 Code) from compacting after every non-empty step
• add a regression test for a 262k context/output model after a 9.5k-token step

Root cause

V2 subtracted the raw advertised output limit from the context limit. For models with output === context, the automatic compaction threshold became zero.

Tests

• bun test test/session-runner.test.ts
• bun typecheck