Release 1.8.3 · Experimental landing-page work, richer media intelligence, and safer org chat streaming

Landing-page routing keeps evolving behind the experimental banner: clearer routing version titles and lighter chrome so the editor stays focused on the graph. Brand media repository picks up intelligence debug and table polish, scroll-stoppers navigation is trimmed, and shared copy routes through the normal translation pipeline.

Media intelligence now treats each asset embedding slot as unique (segment, modality, instruction, slice), replaces artifacts atomically on successful runs, bumps metadata epochs on failure or completion, and exposes an admin-only vector debug search API wired to a small repository debug panel. Migrations dedupe legacy rows, add embedding_slice_index, and backfill repository media kinds where needed; BigQuery ops notes cover dropping legacy media-analysis tables.

Org AI chat is still experimental and access-gated: the dedicated chat service merges env CORS with a fixed localhost baseline, exposes generation headers through CORS, and applies the right headers when writing raw SSE responses. The dashboard transport can replay persisted generation events into UI chunks after refresh instead of leaving reconnect unimplemented.

Control plane surfaces reuse general approvals copy for consistency, and data-table export and view tooling gets usability fixes. Deployment scripts, unified backend Cloud Build, sync-server deploy hooks, and starter GCP monitoring assets move with the stack; internal architecture docs describe the chat service and repository behavior.

These capabilities remain experimental: expect rough edges, partial rollout, and ongoing tuning before we treat them as generally available.