v0.96.0

Fixed

• The CLI no longer aborts with UnicodeEncodeError on a non-UTF-8
  console (Windows cp1250 and friends): the entry point reconfigures
  stdout/stderr to UTF-8, so the arrows, bullets, and sparkline glyphs
  the commands print survive a legacy code page instead of crashing the
  whole command. Found on the first Windows run of synapse doctor;
  harmless where the streams are already UTF-8.

Added

• synapse causality health --since TS — bound the scan to recent
  events on a large log, mirroring the trust graph's --since; a task
  whose entire recorded lifecycle predates the window is not assessed,
  and a window-straddling task is judged on the window's evidence only.
• synapse benchmark --trend --export-csv FILE — the stored history as
  long-format CSV (one row per metric value, context columns on every
  row) for spreadsheets and external monitors.
• The README now points at the cockpit build instructions and documents
  the state snapshot's dead_letters section.
• Log-derived signals reach Prometheus through the node_exporter
  textfile collector: synapse reliability --textfile FILE and
  synapse causality health --textfile FILE write the reliability
  findings and causal-health anomalies as valid labelled exposition
  (synapse_reliability_findings{kind}, synapse_causal_health_anomalies {shape}), deterministic over a given log and parsed back by the
  Prometheus client in the suite so node_exporter never rejects one.
  Two alert rules ship with the observability bundle, and
  docs/observability.md documents the analytics plane landing beside
  the live counters — evidence gauges, not grades.
• The WASM sandbox gained an adversarial proof battery and run
  attestation. tests/test_wasm_sandbox_escapes.py drives a hostile
  module past every limit — memory bomb, fuel bomb, wall-clock runaway,
  a reach for a host syscall, a reach for the network — and asserts each
  is contained by a mechanism (an undefined import cannot link; a grow
  past the cap is refused; the epoch timer interrupts a fuel-free loop),
  not by the tool's good behaviour. synapse sandbox run --attest DB
  appends the run receipt to a durable event store as a sandbox_run
  event, auditable through synapse event-query and replay without the
  tool's bytes ever entering the log. docs/sandbox-threat-model.md
  states what is denied by what mechanism and what is out of scope.
• An observability provisioning bundle under
  integrations/observability/: a Prometheus scrape job, six alerting
  rules over the decision counters (hub down, dead letters growing,
  denials outpacing grants, auth failures, takeover quarantine,
  federation denials), and a committed Grafana dashboard — import,
  pick the datasource, done. A drift-guard test pins every metric name
  the bundle references to the registry, so a renamed metric fails the
  suite instead of silently emptying a panel. docs/observability.md
  walks the five-minute setup and states the plane boundary: /metrics
  is the live process deciding, the store feeds are log analytics.
• The hub's /metrics endpoint grew from 8 to 21 metrics: decision
  counters wired at the decision sites — claims granted/denied, releases,
  directed and broadcast chat, per-message auth failures, rate-limit
  rejections, federation-gate denials, waiter takeovers and their
  quarantines — plus live gauges for connected -rx waiters and the
  dead-letter ledger (targets and letters). Each increment is one integer
  addition in the message path and a scrape stays I/O-free, so an alert
  rule can now see the hub deciding, not just existing.
• /metrics.json dashboard feed — store-attested log metrics for the
  cockpit's metrics panel (total and per-kind event counts, plus the same
  split over trailing hour/day windows), measured against the log's own
  final timestamp so the document is deterministic and replayable; same
  --feeds-db posture as the other store feeds (available with the hub
  down, 404 unconfigured, 503 fail-visible), and the document itself
  states that the live process registry remains the hub's own /metrics
  endpoint.
• synapse doctor --notify-cmd CMD — pipe any warn/fail findings to an
  operator sink command (one line per finding with the remedy attached,
  hub URI in SYNAPSE_DOCTOR_URI), turning diagnostics into proactive
  alerts; a healthy run sends nothing, --fix pages the post-repair
  state, --json composes (stdout stays one document), and the sink is
  best-effort under the same no-shell contract as
  cross-repo --notify-cmd.
• synapse lock --release-timeout SECONDS — tune how long the exit is
  held for the hub's release confirmation on slower links; the wait
  stays bounded either way and the lease TTL remains the backstop.
  Distinct from --wait-timeout, which bounds acquiring the lease.