Experiment with structural metadata filtering on retrieval

[MichaelSchmidle]

Migrated from legacy Fizzy card #73 after Memrok work moved to GitHub Issues + milestones.

Goal

Experiment with structural metadata filtering on top of semantic retrieval.

Why

Stable containers may improve retrieval hygiene. Test whether person / project / topic filters improve precision versus flat embedding search alone.

Requirements

• measure across multiple topics, not only the Memrok dev thread
• compare precision/recall tradeoffs against current retrieval
• keep it framed as an experiment until it earns its way into the hot path

[MichaelSchmidle]

Follow-up note after opening #18: structural metadata filtering may become much cleaner if Memrok first makes local vs global representations explicit instead of treating this purely as a flat retrieval/ranking problem. I would revisit #12 through that lens before implementation.