Orchestrates memory strategy execution for a conversation's message list.

The pipeline:

1. Extracts pinned messages (role: :system or pinned: true)
2. Sorts strategies by priority (lower = runs first)
3. Applies each strategy sequentially on non-pinned messages
4. Re-injects pinned messages at the beginning (in original order)
5. Returns {:ok, filtered_messages}

Presets

Pipeline.preset(:default)     # SlidingWindow last: 50
Pipeline.preset(:aggressive)  # TokenTruncation max_tokens: 4096
Pipeline.preset(:summarize)   # Summarization + SlidingWindow

Note on pinned message ordering: Pinned messages (role: :system or pinned: true) are always placed at the beginning of the output list, regardless of their original position. This is intentional — LLMs expect system instructions at the start of the message list. If you need a pinned message at a specific position, consider injecting it manually after the pipeline runs.