[Hillary Mutisya (hillary-mutisya)]

Summary

• Reduce repeated memory read/index work with committed-generation reads,
  immutable reader/view caches, parallel retrieval and paced background warmup.
• Add opt-in serving experiments and independent query, initial-reader and
  deeper-answer model roles, with cancellation, timing and structured-response
  validation.
• Align Memory Hub Search/Ask and Understanding with the explanation-first
  design while preserving grounding, exact-revision inspection and follow-ons.
• Repair the inherited memory MCP manifest ordering violation so the repository
  policy check passes; dependency versions and manifest semantics are unchanged.

Base integration

Fetched origin/main before publication. Its latest commit, 2ff659b94, was
already an ancestor of this branch; git merge origin/main reported already
up to date. No conflicts or artificial merge commit were needed.

Motivation

Long imported conversations and maintained knowledge views expose expensive
repeated index loading and serial reading work. Users also need a readable
explanation of the evidence, not a page dominated by metadata or visual modules.
This change improves the normal read path and makes performance alternatives
measurable without promoting experimental quality/recall tradeoffs to defaults.

Behavior and configuration

• Foreground reads use available committed generations rather than waiting on
  background ingestion/maintenance. Warm a collection when opened and pace
  background warmup five seconds after completion, with a startup trigger.
• Reuse immutable readers and view indexes with bounded cache ownership and
  invalidation. Retain corpus scope, cancellation and exact-revision semantics.
• P2/P3 serving alternatives remain opt-in through
  TYPEAGENT_MEMORY_PERF_EXPERIMENTS; unknown switches fail explicitly.
  Provisional source previews are not validated answers or follow-on handles.
• Support independently configured query/initial/deeper model roles and Terra
  capacity metadata. Responses API calls preserve structured-output schemas.
  A status/reference disagreement permits one same-model regeneration; other
  invalid responses are not converted into successful answers.
• Search/Ask keeps answer and original evidence in the main column, with typed
  insights in a supporting rail. Understanding adds readable state badges,
  formatted prose, in-flow unknowns, compact source disclosures and a follow-on
  composer directly beneath the explanation. Existing theme tokens remain.
• No attachment/depth UI, new multi-collection guarantees or automatic promotion
  of the experimental fast-reader configuration.

Validation

Verified locally on Windows with Node 24.22.1, at head 63679076a:

• Full pnpm run build passed in 67.583 seconds.
• Browser-client suite: 44 suites / 444 tests passed.
• Additional focused offline check-in regression run: 335 passed, including
  memory service (191), KnowPro (35), AI client (52), agent server (7),
  conversation persistence (20), and gateway/benchmark tooling (30).
• One existing disabled mock-REST host-enablement test remains skipped; the
  focused native model-option/startup tests pass. No new skips were introduced.
• Lint, complexity, circular-dependency and focused/skipped-test gates passed
  against origin/main; formatting passed. Ratchets were run on the committed
  final head, not an incomplete uncommitted file selection.
• Actual product renderers were checked with synthetic RPC responses at
  390/1600 pixels in light/dark themes. Overflow and desktop reading/rail
  geometry checks pass. These fixtures are not live retrieval/latency evidence.
• Repository policy: 12,104 checks passed. The manifest-only repair was also
  checked for parsed-JSON equality before and after the ordering change.