memware 0.8.0

Added

• An optional relevance filter for prompt-time injection, off by default. The prompt hook and
  the Hermes provider's prefetch pick beliefs by keyword, so a prompt about an incident report
  also got a weekly report's file path. With memware config relevance.mode filter, memware asks
  TypeSafe's System One model (jev-1.13.0, pinned) one yes/no question per candidate, all in one
  request, and injects the candidates at or above relevance.threshold, at most k. shadow makes
  the same call, logs each (prompt, candidate) pair to relevance-log.jsonl for calibration, and
  injects what off does. When it is on, the prompt (up to 2,000 characters) and up to 20
  candidate facts go to api.typesafe.ai, and nothing else does. A task notification, a hook
  inside a subagent and a session memware keeps out of its store are never sent. With the filter
  off, memware reads no key, opens no socket and prints the same bytes as before. It fails open
  on a missing key, a timeout, an HTTP error, a redirect or a malformed reply, with one attempt
  under a 1.5 s deadline. With a full pool of 20 candidates the hook took 553 ms p50 and 753 ms
  p95, against 84 ms and 99 ms off. It uses the standard library only, with no new dependency.
  memware config refuses a bad relevance.* value and says what switching the filter on sends.
  memware nuke removes the log and relevance-usage.jsonl.