feat: add /output-limit for a per-model max output budget

[elijahwander]

Summary

• Generalizes the rule engine to target either limit.context (existing /context-limit) or limit.output (new /output-limit).
• /output-limit shares the same parsing and pattern matching: plain tokens, 16K, 1M, 50%, 0 clears, and * / provider/* / exact patterns (longest match wins).
• Both budgets clamp to the field's catalog value, so they only lower, never raise.
• Context and output rules are stored separately (context-limit / output-limit); one catalog transform applies both.

Motivation

A per-model max output budget caps reply length/cost, complementing the existing context budget that moves the compaction point.

Test plan

• bun test — 20 pass, 0 fail (added applyBudget output lowering + ceiling clamp, and /output-limit set/separate-storage/display tests).
• No imports added; plugin remains a dependency-free plain { id, setup } object.

Notes

• Bumped VERSION and package.json to 0.2.0.
• Kept the existing ## OpenCode referral section unchanged.