v0.1.34

Upgrade notes

• Background LTM warm-up is enabled by default (#664, #671): server and
  daemon startup now launches a best-effort, host-owned LTM warm-up task so
  the usual ~9s child/model cold start is paid before the first surfacing
  call. It never blocks the proxy MCP initialize handshake, and lazy start
  remains the retry path after failure. Set
  MEMTOMEM_STM_SURFACING__WARMUP_ENABLED=false when spawning one LTM child
  per proxy process is undesirable, such as with many short-lived proxies.

Added

• Background LTM warm-up at server/daemon startup, controlled by the new
  surfacing.warmup_enabled setting (default: true). (#664, #671)
• Durable surfacing fault counters (#666): timeout / breaker-open / degraded-LTM
  skips now persist day-aggregated to surfacing_faults in the feedback DB,
  and mms stats renders them with a warning line — previously these signals
  lived only in per-process memory (stm_surfacing_stats), so a surfacing
  pipeline dead on LTM timeouts for days looked merely quiet from the CLI.
  Rows share the stats_retention_days sweep.

Security

• Surfaced memory fields are now serialized as inert, single-line Markdown
  data before injection. Control/bidi characters, HTML delimiters/entities,
  Markdown delimiters, and Unicode compatibility confusables can no longer
  create a second <surfaced-memories> boundary. Invalid memory ids omit only
  the copyable id token; feedback for the surfacing remains available.
  (GHSA-43hx-xm7w-3mhj)
• LTM recovery is generation-aware and shares exactly one reconnect flight per
  dirty session. Caller cancellation no longer aborts a reconnect for other
  waiters, and cancellation during any post-reconnect retry dirties only that
  new session generation.
  (GHSA-72jh-722p-rqr3)
• Raised runtime and optional-extra dependency floors to patched versions.
  Locked runtime and extras audits are now blocking CI and release gates;
  the full development closure remains advisory. (#673)