fix(subagent): raise maintenance delegate token ceilings for dense archives

[Shimmernight]

Problem

When runtime MEMORY.md is at capacity, every mnemon_runtime_memory add delegates to maintenance workers (archive routing / local compaction / metadata sampling). Those workers are capped at 8192 (and 4096 for metadata-maintenance) completion tokens. Reasoning models spend most of that budget on analysis before emitting the small structured result, so the worker stops with stopReason=max-tokens and the whole tool call aborts — with zero writes.

Reproducible with dense CJK content. Reported upstream as #70 (still fails on 0.3.0) and #83/#85 (same symptom, auto-closed for form issues).

Root cause

src/subagent.ts — static per-op ceilings override the host-resolved token budget:

const perOpMaxTokens = operation === 'migration' || operation === 'compaction' || operation === 'document-archive' ? 8_192
  : operation === 'metadata-maintenance' ? 4_096
  : undefined

Fix

Raise the ceilings to 32768 / 16384. The structured output of these workers is still small; the caps only bounded the analysis budget, which is the part that needs headroom. Fixes-class change: no new config surface, no issue gate required.

• src/subagent.ts: 8_192 → 32_768, 4_096 → 16_384
• tests/subagent.spec.ts: update the pinned expectations (metadata 4_096 → 16_384; migration/compaction/merge 8_192 → 32_768)

Verification

• pnpm exec tsc --noEmit — clean
• pnpm exec vitest run tests/subagent.spec.ts — 42/42 passing

Notes

The static ceilings remain finite. #70's deterministic bulk-import direction would remove the LLM dependency entirely; this PR is a targeted stopgap until that lands.

[Grivn]

感谢提交。最新 main 已通过 #102 支持配置 Runtime 迁移和压缩的 token 预算，容量归档也已有确定性写入及模型路由失败回退，因此不再采用统一提高固定上限的方案，本次关闭此 PR。

文档归档和元数据维护仍使用固定预算。如果这些路径在最新 main 上仍出现 max-tokens，可以提供对应复现，再针对具体场景调整。