[Feature]: (memory): optional PowerMem retrieval for system prompt context

What workflow is missing or painful today?

Motivation

OpenHarness memory today is Markdown-first: MEMORY.md plus topic files under the project memory directory, with keyword/heuristic relevance via find_relevant_memories. Teams using PowerMem (vector + full-text + graph, optional intelligent memory features) have no first-class way to inject retrieved memories into the harness system prompt.

Proposed change

Add an optional integration layer that:

• Keeps default behavior unchanged (memory.backend: local).
• Supports HTTP against a running PowerMem API server (POST /api/v1/memories/search).
• Optionally supports in-process SDK via optional extra openharness-ai[powermem] (requires Python compatible with the powermem package, typically 3.11+).
• Supports hybrid: existing local Markdown relevance plus PowerMem retrieval.

Alternatives considered

No response

Additional context

No response