v2.1.0-rc.1

Security

• Dashboard marker checks now describe only marker presence. Strict embed authorization requires
  an explicit tenant-to-dashboard assignment and trusted isolation evidence bound to the served
  revision and definition digest. Existing constructors remain available, but strict verification
  fails closed without the new evidence and assignment providers. Migrate strict integrations using
  ADR 0028.
• The Signalboard demo now validates antiforgery tokens on sign-in and sign-out.

Fixed

• External JSON exports parse root arrays and stream complete, ordered chunk chains. Incomplete,
  truncated, malformed, and inconsistent results fail explicitly; expired external links can be
  renewed before rows from that chunk are emitted. Inline materialization has a 25 MiB limit.
• Statement and Genie deadlines include credentials, submission, polling, response bodies, and
  export reads. Accepted Genie conversation ownership is recorded before polling; recoverable
  identifiers remain available when the answer times out.
• Operation dispatch uses a bounded worker pool so a held upstream poll does not serialize all
  tenants. Refresh status treats USER_CANCELED as cancellation and keeps unknown-state errors safe.
• Candidate consumer validation rejects missing or mismatched packages instead of falling back
  to a public or cached LakeWright package. CI fails when advisory scan evidence is unavailable,
  and package compatibility validation again checks the current stable 2.x baseline.
• Signalboard's Docker build includes its four linked public images. Authenticated navigation
  reflows at narrow widths, and starting an operation provides visible status feedback.

Added

• Raw-data columns can opt into exact BIGINT and DECIMAL(p,s) filters through WholeNumber
  and FixedPoint. Existing Number filters retain approximate DOUBLE semantics. See
  numeric contracts for configuration and precision limits.
• Load verification records bounded admission, completion, failure, drop, and censoring counts,
  and gates generated throughput separately from completion ratio. Local network and browser
  lab profiles record their conditions without claiming production capacity or field web vitals.