feat: skip assistant boilerplate as recall-primary episodes

[sdntsng]

Summary

Wave D: assistant conversation_episode rows drown user facts (gold “4 weeks” vs congratulations “3 weeks”). LME single-session-assistant is 0/2, but the indicated fix is not “extract more assistant chit-chat” — it is to stop persisting boilerplate as recall-primary episodes while still keeping assistant-stated facts.

Changes

• Deterministic extract: skip phatic assistant turns; store non-phatic assistant utterances as assistant_stated facts (not PrimitiveEpisode).
• Provider extract: same filter after merge so async worker cannot reintroduce congratulations episodes.
• User free-dialogue episodes unchanged.
• Tests: congratulations dropped; FCC/distillation assistant fact retained without conversation_episode.

Test plan

• go test ./internal/memory/
• OpMem 13/13 + marketing after merge (extract filter is conversation-path)
• Do not treat as LME quality lift until a dedicated pin

Linear / tracking

Wave D inverted PR9 after #100.

Risk / follow-ups

• Evidence plane still captures assistant actor_role.
• Phatic cue list is product-general (congratulations / you're welcome / …), not LoCoMo/LME-named.
• No SOTA claim. dev only; do not merge to main.