Namespace-scoped reads: namespace param on memory.recall + feedback event scoping (+ knowledge.compose)

[ohdearquant]

Problem

Benchmark/evaluation harnesses need a frozen, isolated store per measurement arm: writes go to a dedicated namespace so arm N cannot be contaminated by arm N-1 or by live activity. The write side already supports this (memory.remember accepts namespace), but the read side does not:

• memory.recall has no namespace param — a per-arm namespace can be written but not read back in isolation.
• brain.auto_feedback cannot scope its event away from the live posteriors, so serving feedback from a bench arm trains the production brain.
• knowledge.compose likewise has no namespace filter (lower priority).

Verified against the live verb surface 2026-07-08: none of the three accept namespace.

Ask (preference order)

1. namespace param on memory.recall (filter recall to one namespace) and on brain.auto_feedback (scope the emitted event). This alone unblocks the primary use case.
2. Same for knowledge.compose if cheap.
3. Snapshot-pinned recall/compose (Snapshot-pinned reads: optional as-of/snapshot parameter on memory.recall + knowledge.compose #726) remains the ideal end state; this is the pragmatic subset.

Notes

• ADR-007 Rev 6: multi-record ops default to WHERE namespace='local' with an explicit namespace= escape — recall is a multi-record read, so a verb-surface namespace arg aligns with the existing model rather than extending it.
• Feedback scoping needs a design check: whether the event row's namespace stamp is enough to keep bench events out of live posterior updates, or whether the brain update loop needs a namespace filter too.