v3.7.0

Added

• MCP client-sampling synthesis (dark flag MEMO_SAMPLING_SYNTH_ENABLED):
  memo_ask, memo_chat_ask, memo_reflect, memo_synthesize_run, and
  memo_consolidate delegate synthesis to the connected client's model via
  MCP sampling, with sticky per-request MLX fallback, a per-request call cap
  (MEMO_SAMPLING_MAX_CALLS), and a synthesizer attribution field on
  dict-shaped responses. Grounding judgement stays local by design.
  Companion flags: MEMO_SAMPLING_TIMEOUT_S, MEMO_SAMPLING_MAX_TOKENS.
• Chunk emission at save/update time (MEMO_CHUNK_INGEST): long documents
  get their heading-aware chunk records immediately on save()/update()
  instead of waiting for the next manual memo reindex. Best-effort — a
  chunk-emission failure never fails the write; metadata-only updates skip
  emission. Docs for the chunker refreshed to match its real wired state.

Fixed

• Release metadata realigned for v3.6.0: rebuilt .mcpb archives,
  regenerated uv.lock, SECURITY.md supported line bumped to 3.6.x.