v0.4.0

memtomem-stm 0.4.0 — a minor, not a patch. Sixteen behavior changes across the
advertised tool surface, config validation, client import/prune, daemon
concurrency and surfacing telemetry. Two of them break a working setup with no
user action, which is what promotes this release rather than the size of the
diff: a config with +inf as chars_per_token now fails validation at load
(#977), and under the strict exposure profile a tool whose title or icon URL
is credential-shaped is now withheld rather than advertised (#895).

Read the upgrade notes below first. The full entry-by-entry list is in
CHANGELOG.md
— at 50 KB the section is past what a release body should carry, so this note
keeps the aggregate rather than every bullet.

uv tool install --force memtomem-stm     # or: pip install -U memtomem-stm

Upgrade notes

• Proxied tools can look different in your client's picker, and strict can
  withhold more of them (#895, #922). A tool whose upstream sets a top-level
  title now displays as [{server}] {title} instead of its prefixed name, and
  its icons reach the client. A tool that supplies no description now
  advertises its own prefixed name rather than a bare [proxied] prefix. Under
  the strict exposure profile, a tool whose title or icon URL is
  credential-shaped is withheld (sensitive_metadata); under review it is
  demoted. If a tool disappears from your picker after upgrading, check its
  upstream metadata against the credential scan before assuming a regression.

• Two config values that used to load now fail validation at load (#977,
  #929). +inf as chars_per_token at any of the three levels, and a
  max_result_tokens above the signed-64-bit ceiling, are refused at load
  instead of loading and then failing the calls that used them. Separately, a
  per-tool chars_per_token now applies to a token budget inherited from the
  server, so a deployment that sets both sees that tool's resolved char budget
  re-scaled (a server at 400 tokens x 2.5 with a tool ratio of 4.0 moves from
  1000 to 1600 chars).

• mms prune and mms add --from-clients now compare identity, not just
  name (#981, #983, #984, #955). A prune no longer removes a source
  entry whose env or headers diverge from the STM upstream — divergence in a
  second client is reported as a conflict, in the first by name. A prune that
  previously removed a same-name entry from a second client now removes only the
  sources that hold the same server and says which ones it left. Cursor's MCP
  configs are now scanned, so --all adopts more servers than before on a
  machine that has them; Cursor entries are reported but never written.

• The daemon runs several LTM operations at once, and its telemetry says so
  (#874, #994). Set daemon.max_concurrent_ltm_ops to 1 to restore the older
  serialized daemon — that is the only value that does, since a core with no
  parallel capacity degrades the same way at 2 as at 4. mms daemon status
  reports queue.in_flight above 1 and gains queue.concurrency. Latency
  classification changed with it: how a call ended now outranks whether the LTM
  was warm, so a request that connected, searched and then failed is filed as
  the failure it was rather than in cold_samples.

• Durable surfacing telemetry no longer lands before the response (#996).
  Fault counters, diagnostics and retention sweeps are queued when their branch
  runs and land whenever the FIFO worker gets to them. A request cancelled by
  its deadline mid-write still lands its event row, so a shed request can leave
  a row for a manifest its client never received. Reads of stm_surfacing_stats
  immediately after a call may not see the row yet.

• max_upstream_bytes is enforced on the inbound envelope, and the embedding
  scorer falls back further (issue #957, #873). The cap measures the decoded
  message's compact serialization, which is what docs/configuration.md has
  always described; a result that serializes larger than the envelope that
  carried it is no longer rejected on that parsed size. And an embedding
  provider that returns a different number of vectors than it was given inputs
  now falls back to BM25 for that call.