MemoRizz v0.8.0

Added

• Added ordered, bounded StreamEvent delivery through run_stream_events()
  and arun_stream_events(), including answer digests, cancellation, explicit
  terminal status, provider usage, and separate answer/persistence completion.
• Added default incremental answer delivery to the CLI, REPL, Playground and
  its desktop webview, plus MCP progress with an opt-in versioned answer-event
  extension. Added flushed CLI text/JSONL output and --no-stream compatibility.
• Added concrete provider streaming adapters, including native OpenAI Responses
  and Azure Chat streaming, with interruption, backpressure and incomplete-stream
  checks across supported providers.
• Added typed host observability events, trace propagation, memory-selection
  ledgers, artifact lineage, output contracts, browser-acknowledgement evidence,
  coverage profiles, and conservative outcome diagnostics.
• Added native SQLite, MongoDB and Oracle observability indexes, explicit
  migrations, scoped pagination, retention, backfill/parity checks, and operator
  tooling. Added an isolated local Oracle setup and live database test fixtures.
• Added Incident Finder, causal waterfall, memory/artifact/contract inspectors,
  selected-event navigation, structural comparison, health views and responsive
  mobile layouts, with synthetic browser and native WebKit regression fixtures.

Fixed

• Historical bundled traces now retain child events, tool names, timestamps and
  model metadata. Health and insights share authorized selection/count semantics.
• Read completeness no longer implies end-to-end instrumentation or verified
  outcomes. Missing host hooks are shown as unavailable instead of usable actions.
• Search and cross-view links retain trace selection. Legacy JSON-string replay
  source lists no longer become individual-character resource identifiers.
• Streaming failures and cancellation preserve partial text without adding
  exception prose to answers; dumb-terminal REPL output now flushes incrementally.
• Oracle observability initialization now handles its first-install dynamic SQL
  correctly. Audit reports are kept outside the repository and release archives.

Security and compatibility

• Added tenant-scoped, fail-closed operator access, audited explicit reveals,
  transient account/artifact resolver hooks, immutable event envelopes and inert
  replay drafts. MCP observability queries remain opt-in and read-scoped.
• MemAgent.run() still returns a complete string. Whole-answer validators
  continue to buffer until acceptance; streaming does not expose private tool
  drafts or rejected answers. Third-party cancellation remains cooperative.
• Host instrumentation, production migration/backfill and signed desktop
  distribution require separate deployment acceptance. This release does not
  change OpenSpeech or invent missing historical evidence.