v0.89.0

Added

• synapse benchmark measures the installed package on the operator's
  machine: probes for durable event-store appends, journal replay, lite
  relay encoding, and who plus claim-to-grant round-trips over a real
  loopback WebSocket hub, each reporting throughput and p50/p95 latency.
  The scorecard carries the host context — package version, interpreter,
  CPU model and governor, load averages before and after — and an explicit
  shared-workstation isolation label, so the numbers read as regression
  evidence, not as isolated-core production claims. --probe selects a
  subset, --iterations overrides defaults, --json emits data, and
  --results FILE writes the scorecard to disk.
• synapse cross-repo widens coordination from one repository to a whole
  checkout tree: it scans every repository under a root directory for
  dependency manifests (pyproject.toml, Cargo.toml, package.json,
  go.mod) and CODEOWNERS files, composes them into a graph of dependency
  and shared_owner edges, and joins the live claims of a hub event log onto
  it (a claim's worktree is its repository). With --repo the exit code
  becomes a coordination signal — 1 when a live claim exists in a
  repository connected to the focus by a dependency edge — and --json and
  --dot emit the graph as data or a Graphviz digraph. Manifests that exist
  but cannot be parsed are reported as problems rather than silently
  skipped. Declaration-level, advisory evidence only.
• synapse trust-graph queries the durable event log as an evidence graph,
  realising the agent-trust-graph design's read-only projection: typed edges
  between agent and task nodes — positive release receipts, stale claims,
  declared failed checks, broken handoff candidates, and one agent-to-agent
  edge per reconstructed conflict pair — each carrying the event-log sequence,
  timestamp, and evidence fields that created it. --agent, --task, and
  --since (a decay window) focus a review; --json emits the graph as data
  and --dot as a Graphviz digraph. Evidence with provenance, not scores: no
  ranking, no grades, no authorisation.
• HubConfig groups the forty-odd SynapseHub keyword parameters into typed,
  frozen family records — HubLimits (every enforced ceiling),
  TakeoverDamping, HubAuthConfig (connection, per-message, and ACL
  enforcement), HubMetricsConfig, MultiHubConfig, and FederationConfig —
  and SynapseHub.from_config(config) builds a hub from the record.
  Behaviour is identical by construction: every field name and default
  mirrors its keyword parameter, contract tests pin the flattened record
  against the live signature so the two surfaces cannot drift, and the flat
  keyword surface and every CLI flag remain unchanged.

Changed

• import synapse_channel now resolves its public names lazily (PEP 562):
  the submodule behind a name is imported on first attribute access, cutting
  the bare package import from roughly one second to under ten milliseconds
  while keeping __all__, every re-exported object, and type-checker
  visibility identical.
• The synapse CLI registers subcommands lazily: main reads the requested
  command off argv and imports only the module family that owns it, so a
  short call such as synapse who or synapse merkle root no longer pays
  the import cost of the whole surface (local commands start in roughly a
  quarter of the previous time). --help, --version, and unknown commands
  still build the full parser, and contract tests pin every registration
  unit to the exact commands it provides and to help output identical to the
  full build.

Added

• synapse status --watch refreshes the one-line hub summary every
  --interval seconds (default 2) as an operator dashboard. Each refresh
  opens its own probe connection, so a hub restart shows as an honest offline
  line; a TTY rewrites the line in place while piped output appends one line
  per refresh, and --json --watch streams one JSON object per line (NDJSON).
  --count N bounds the refreshes; Ctrl-C stops an unbounded watch cleanly
  with exit 0, and the bounded form exits with the last observed state.
• synapse workflow contention joins a declarative workflow to the durable
  log: it compiles the workflow to its task ids, runs the same offline
  yield-advice analysis as synapse causality contention, and keeps only the
  overlapping live-claim pairs a workflow task is party to — whether it keeps
  or yields. Pairs outside the workflow are counted in a trailing note; the
  exit code signals scoped collisions only (0 none, 1 at least one, 2 on
  an invalid workflow, a missing store, or the node ceiling).
• synapse participant convene --dry-run prints the convocation plan without
  taking a single turn: the resolved mode, its round count, and each seat's
  identity, readiness, planned turns, and estimated cost from an
  operator-supplied --pricing table under printed per-turn token assumptions
  (--est-input-tokens/--est-output-tokens). Seats without a price line are
  reported unpriced and excluded from the total; with --budget-usd the report
  states whether the estimate fits. Exit 0 when every seat is ready, 1 when
  any is unavailable, 2 for a refused configuration.
• The repository root now ships a composite GitHub Action (action.yml)
  wrapping synapse policy-check, so a repository can gate CI on a release
  receipt — optionally recomputing the Merkle commitment and requiring a
  trusted hub signature — with a single uses: step. Inputs reach the shell
  through environment variables, never script interpolation; the decision
  report is exposed as the report step output.
• The synapse causality contention documentation gained a worked two-agent
  example with real command output, showing how downstream weight picks the
  yielder and how a tie falls back to first-come precedence.
• synapse participant costs reads opt-in session telemetry back from a hub
  SQLite event store — offline, like synapse accounting report — and prints
  the latest cumulative snapshot per (agent, session) (turns, errors,
  abstentions, token pressure, metered spend, mean latency, highest rate-limit
  utilisation seen) plus fleet totals, or the machine-readable report with
  --json. Where the accounting report answers what models cost, this answers
  how participant sessions are going and what they spent; a missing store
  refuses with exit 2.