design: brain posterior read-path is unwired — recall ranking never consumes profiles (BrainProfileHint + PackTunable dead)

[ohdearquant]

Calibration study 2026-06-10 (workspace .khive/workspaces/20260610/recall-calibration/memo.md) confirmed what #62 suspected and Ocean observed directly ("I don't see how brain profile is impacting the retrieval"):

Verified by call graph (grep + read, spot-checked independently):

• PackTunable::project_config() / apply_config() (khive-brain-core trait, khive-pack-memory impl) have zero production callers — only the impl's own unit tests call them.
• BrainProfileHint in RecallConfig (crates/khive-pack-memory/src/config.rs:83) is dead config — defined, serde-defaulted, never read at ranking time.
• The write path works (brain.feedback → events → posteriors). The read path (resolve profile → project posteriors into scoring config → apply per-recall) does not exist.

Needed: a design decision (ADR-level — it changes the pack contract) for where posterior projection happens:

1. Per-recall: recall handler resolves caller's bound profile (brain.resolve semantics) and projects posteriors into the scoring config before ranking — fresh but adds a lookup per recall.
2. On-activation: brain.activate pushes projected config into the memory pack's tunable state; recall reads cached state — fast, but staleness between updates.
3. Hybrid: cached projection invalidated by brain event epoch.

Whichever is chosen must define: which posteriors map to which scoring tunables (the calibration variable inventory lists 22), the fallback when no profile is bound (global defaults), and namespace/actor resolution order (relates #75, and feedback auto-resolution per #62).

Relates #62 (posterior-serving verification is the test this design must make passable). Blocks any "profile impact" claims in product surface.