Revert "feat: add 4 scorer kinds (retrieval, demotion, procedure, dedup) + dashboard surface + dedup React install"

[rahuldindor411-glitch]

Reverts #40## Intent

AgentProbe's existing memory evaluation routes everything through an
LLM-judge over a transcript. That is the right shape for conversational
behaviors, but it cannot answer the question the AutoGPT dream system
needs answered: given a query, did the right memories come back in
the right order. This PR adds the missing scorer kind and the
scenarios that exercise it.

It ships:

1. A pure ranking-metrics module covering precision@k, recall@k, MRR,
   and NDCG@k (with log2(rank + 1) discount), pinned by 25 known-answer
   unit tests.
2. A new retrieval: block on scenario YAML — golden, optional
   forbidden, k, match, per-metric weight, pass_threshold,
   and a source (either raw_exchange_key for inline adapter
   payloads or fixture for offline runs).
3. Runner integration that scores retrieval alongside the LLM judge,
   forces a scenario fail when any forbidden item lands in the top-k,
   and surfaces a RetrievalScore on ScenarioRunResult.
4. A new retrieval_scores persistence table mirroring
   judge_dimension_scores on both SQLite (v8 → v9) and Postgres
   (v4 → v5), with recorders, getters, and Drizzle schema entries.
5. Report-renderer support — retrieval_rows + a pretty JSON blob
   alongside the existing judge dimension rows.
6. Five new memory scenarios (data/retrieval-memory.yaml + fixtures
   under data/fixtures/retrieval/) that exercise forget-on-request,
   warm-context relevance, stale-fact demotion, scope filtering, and
   bounded cascading expiry.

Behavior changes

• Scenarios may now declare a retrieval: block. Without one, behavior
  is unchanged: the judge is still the only signal, the overall pass
  is whatever the judge said.
• With a retrieval: block present, the runner scores the retrieved
  list (resolved from last_reply.rawExchange[<key>] or a JSON fixture)
  and the overall scenario passed becomes judge.passed && retrieval.passed.
  A forbidden-item hit in the top-k forces retrieval.passed = false
  regardless of the weighted aggregate.
• Persistence schema bumps: SQLite SCHEMA_VERSION 8 → 9, Postgres
  POSTGRES_TARGET_VERSION 4 → 5. Both backends ship the new
  retrieval_scores table in their baseline DDL and add a forward-only
  migration step that create table if not existss it on upgrade.
• Run-history readers now return retrievalScores: Array<...> on each
  scenario record. The field is optional on the ScenarioRecord type
  so existing tools that don't read it stay typecheck-clean.

How it's wired

• Math: src/domains/evaluation/ranking.ts is pure TypeScript with
  no I/O and 100% function coverage from
  src/domains/evaluation/ranking.test.ts. The four metric primitives
  are individually exported (and individually tested against
  known-answer cases like NDCG@3 of [1, 0, 1] and MRR of [0, 0, 1]),
  and a top-level scoreRanking aggregator threads them together with
  a weighted average, forbidden-hit detection, and a pass threshold.
• Scorer: src/domains/evaluation/retrieval-scorer.ts consumes
  the parsed YAML config plus a RetrievalSourceContext
  (scenariosPath, lastAdapterReply) and resolves the actual
  retrieved list. It accepts strings, {label}/{text}/{name}/{id}
  object payloads, or a JSON file path. Missing sources record a
  source: "missing" score instead of crashing the suite.
• Schema: src/domains/validation/load-suite.ts validates the
  retrieval: block strictly — non-empty golden, valid match policy,
  threshold ∈ [0,1], known metric weight keys, non-negative weights.
• Runner: runScenario in src/domains/evaluation/run-suite.ts
  scores retrieval after the judge phase, records it via a new
  recordRetrievalResult recorder hook, and ANDs the two pass flags
  into the final scenario result.
• Persistence: both backends get an idempotent
  create table if not exists retrieval_scores in the baseline DDL,
  a forward migration step, an ensureRetrievalScoresTable helper for
  the legacy SQLite in-recorder migrator, recorder methods, and a
  getRun join. The Drizzle inventory test pins the table count.
• Reporting: prepareScenarioView emits retrieval_rows,
  one row per metric with pretty value/percent labels, and a
  retrieval_scores_pretty JSON blob. Templates that already exist
  see no change until they reference the new keys.

Scope rails

• No new env vars. No new top-level deps.
• Adapter abstractions untouched. The scorer reads rawExchange
  data the adapter already returns; no EndpointAdapter shape change.
  Scenarios that need offline scoring use the fixture source.
• No frontend dashboard work. The data is persisted and surfaced
  in the report-prepared view, but the React dashboard is not
  updated. Flagged as a follow-up below.
• No mass-rewrite of existing scenarios. The five new scenarios
  live in their own data/retrieval-memory.yaml pack alongside the
  existing data/multi-session-memory.yaml. The forget-on-request
  scenario is a ranking-scored sibling of the existing judge-only
  mem-negative-forget-on-request.

Risk

• Persistence migration touches both backends. The new table is
  additive and the migration is create table if not exists, so the
  forward path is safe. Rollback to prior schema would require
  dropping the table.
• Scenarios with a retrieval: block but no fixture and no adapter
  payload will score source: "missing" and fail. This is intentional
  — silently passing a "we got nothing" run would mask real regressions
  — but it does mean integrating this against an adapter that doesn't
  populate rawExchange.retrieved requires that adapter to be updated
  or the scenarios to use the fixture source.
• The RetrievalScore is ANDed with the judge for the final pass flag.
  A scenario that previously passed via judge alone and now adds a
  retrieval block that fails will see its overall pass flip. That is
  the intended UX — both signals must be green — but it's a behavior
  change worth calling out.

Out of scope (deferred)

• Dashboard UI for retrieval scores. The data is persisted and
  prepared for the renderer; React surfaces are not updated. The
  retrieval_rows array is structured the same way as dimension_rows
  so a UI addition should be a small follow-up.
• Adapter-side retrieval emission. Wiring AutoGPT or OpenClaw to
  return rawExchange.retrieved payloads is left as adapter work and
  doesn't belong in this scorer-shaped PR. The scenarios ship with
  fixtures so the scorer math is exercised against curated graph
  snapshots in the meantime.
• Multi-judge aggregation. AgentProbe's bias_mitigation.multiple_judges
  config remains unimplemented (preexisting); not in scope here.
• Sweeping the remaining 24 unused multi-session-memory scenarios
  into the new format. Out of scope; the brief calls for ≥5 new
  scenarios and that's what shipped.

Validation

• bun test: 250 pass / 7 fail. The 7 failures are all
  pre-existing dashboard React tests
  (useDashboard, RunDetailView SSE, StartRunView, CompareView)
  that fail identically on origin/main at the parent commit. They
  are not touched by this PR.
• bun run lint: clean.
• bunx tsc --noEmit (root): clean.
• bun run docs:validate: clean (new scenario added to platform.md,
  current-state.md, e2e-checklist.md; generated docs refreshed).
• Persistence migration: SQLite baseline + forward migration both
  exercised by tests/unit/db.test.ts and
  tests/unit/persistence/migrations.test.ts. Postgres migration
  exercised by tests/unit/persistence/migrations.test.ts against a
  pre-v5 baseline.
• Scenario pack: every scenario parses, references a known memory
  rubric, and points at a fixture file that exists relative to the
  YAML. Pinned by tests/unit/retrieval-memory.test.ts.

Checklist

• bun run lint clean.
• bunx tsc --noEmit (root) clean.
• bun run docs:validate clean.
• bun test shows no new failures vs. origin/main.
• No new env vars.
• No secrets committed.
• Behavior docs updated (docs/product-specs/platform.md +
  current-state.md + e2e-checklist.md).

Commits

• feat(evaluation): add ranking metrics primitives — pure math module + 25 unit tests
• feat(evaluation): add retrieval scorer kind and YAML schema — typed config, scorer module, loader, schema validation
• feat(evaluation): wire retrieval scorer into runScenario — runner integration + two runner-level tests
• feat(persistence): persist retrieval scores in sqlite and postgres — table, migrations, recorders, getters
• feat(reporting): expose retrieval metrics in run reports — prepared view fields + biome touch-ups
• feat(data): add five ranking-scored memory scenarios — YAML pack + fixtures + pack-level tests
• docs(specs): document ranking-scored retrieval scenarios — platform spec entry + generated docs refresh

[rahuldindor411-glitch]

https://www.instagram.com/rekha__kumari__dindor?igsh=emJocXgxZW9pNGxx

[rahuldindor411-glitch]

Rekha dindor

[rahuldindor411-glitch]

1. 8209423503