Add deterministic overflow archive strategy for compaction

[chubes4]

Problem

Some inputs are too large to safely send through a model summarizer. Data Machine hit this with a ~250KB MEMORY.md: the request reached OpenAI but returned 0 bytes until the full request timeout. Raising timeouts is not the right fix.

Goal

Agents API should provide a deterministic overflow strategy that preserves content without invoking a model.

Proposed Shape

• Policy fields:
  • overflow threshold
  • target retained size/count
  • archive strategy enabled/disabled
• Strategy behavior:
  • split ordered items at safe boundaries
  • retain a small active subset
  • archive the remainder verbatim
  • emit pointer/stub metadata for where archived content should be referenced
  • never discard original content
• The consumer supplies persistence/archive destinations; Agents API only returns the transformed item sets and metadata.

Acceptance Criteria

• Oversized inputs can be transformed without calling a summarizer.
• Archived items are returned verbatim with IDs and metadata intact.
• Retained output includes enough metadata for a consumer to write an archive pointer.
• Tests cover large input, no-op small input, single unsplittable item, and deterministic output.

Notes

This generalizes Data Machine's runaway Daily Memory circuit breaker without moving Data Machine file writes into Agents API.