memd v0.60.0

Added

• Added write-admission guardrails for public memory writes so obvious
  low-signal progress chatter and generated digest wrappers do not become
  durable memory by default, while explicit priority / importance tags and
  concrete evidence, decisions, commands, paths, fixes, and validation results
  remain durable.
• Added memd eval-retrieval and memd eval-write-quality gates for checking
  retrieval quality, known-useful recall, admission behavior, dedupe behavior,
  storage growth, and generated-digest suppression with machine-readable
  reports.
• Added memd audit, memd cleanup-plan, memd purge, and
  memd purge-archive operations for measuring memory bloat, producing
  approval-first cleanup plans, archiving purge candidates, and verifying purge
  archives.
• Added operational documentation for useful memory writes, cleanup safety,
  archive-first purge workflows, and memd-vs-alternative positioning.

Changed

• memd memory-md, memd search, and session-start context now bias more
  strongly toward durable, project-scoped, non-generated lessons and suppress
  low-value generated digest noise unless explicitly requested.
• Cleanup plans now include estimated purge batches plus exact destructive and
  archive-verification command previews so large stores can be cleaned in
  reviewable chunks.
• memd purge --apply now verifies the written archive before mutating the
  store and reports the verification result in the completed purge response.

Verified

• Re-ran the release BEIR retrieval benchmark gate with the CI-pinned fiqa +
  scidocs parameters. Candidate normalized nDCG@10 was 0.571, and the paired
  regression gate passed against evals/bench/baselines/beir_v1.json
  (candidate_mean=0.536, baseline_mean=0.417, 47 paired queries).