claude_memory 0.9.1 → 0.11.0

data/docs/quality_review.md

@@ -1,285 +1,637 @@
 # Code Quality Review - Ruby Best Practices
 
-**Review Date:** 2026-03-19
-**Previous Review:** 2026-03-09
-**Last Quality Update:** 2026-03-20 (13 items completed)
+**Review Date:** 2026-04-28
+**Previous Review:** 2026-04-22 (6 days ago)
+**Last Quality Update:** 2026-04-22 (4 items completed — LLMCache + MetricsAggregator extractions, Publish DRY, Dashboard specs)
+**Codebase Growth:** 17,014 → 19,025 LOC (+2,011, +12% in 6 days)
+
+> **Post-review update (2026-04-28, same session):** Items #34, #32, #35, #36 (quick wins + missing command specs) and the first two of the six proposed `Dashboard::API` extractions (#31: `Dashboard::Timeline` + `Dashboard::Health`) landed before the v0.10.0 release commit. `dashboard/api.rb` dropped 807 → 607 LOC (-200, -25%), reversing the regression and bringing it back under the 2026-04-22 baseline (627). The remaining four extractions (`RecallQuery`, `RecallTriggerFinder`, `UserPromptExtractor`, `FactsQuery`) are deferred to 0.10.1. Findings below describe the *pre-update* state captured at review time.
+
+---
+
+## Post-0.11 Investigation: Hallucination Rate Metric Calibration (2026-04-30)
+
+When #48 (hallucination-rate metric) was first run against this project's real DB, it surfaced numbers that *looked* alarming:
+
+- Quality score: 39/100
+- Bare conclusions: 34 / 59 active facts (57.6%)
+- 7-day rejection rate: 27 of 32 facts (84.4%)
+
+The first read was that the LLM extractor was producing noise faster than usable knowledge. Per `improvements.md` #60, four causes were proposed; diagnostics ran 2026-04-30:
+
+| Cause | Verdict | Evidence |
+|---|---|---|
+| Prompt drift in `distill-transcripts.md` | **Confirmed dominant** | 34/35 (97%) bare-conclusion facts pre-date the reason-clause prompt commit `f22d12f` (2026-04-20). Only 1 was created post-commit (and that one is a meta-convention added during this session). |
+| Auto-memory mirror regurgitation | Rejected | 0/35 substring matches in `~/.claude/projects/.../memory/*.md`. Auto-memory mirror only landed in 0.10.0 (2026-04-28), after the bare-fact creation window — temporally impossible to be the source. |
+| `ReferenceMaterialDetector` predicate scope too narrow | Not material | Only 3/35 bare facts are `decision`-predicate; 0 of those match the strong reference-material patterns. Expanding `GUARDED_PREDICATES` would not move the needle on the bare-conclusion count. |
+| Junky corpus / rejection cluster | **Confirmed in single class** | All 27 rejected facts in the 7-day window are `uses_database` (18) or `deployment_platform` (9), all with `session_id=nil` (MCP-originated, almost certainly `/study-repo` runs misattributing external-project tech to this project), all from 2026-04-23 to 04-24. Systemic single-class failure, correctly cleaned up after detection — not ongoing extraction noise. |
+
+**What this means for #48 as currently shipped:**
+
+The metric is *technically correct* but *pragmatically misleading*. It bakes historical noise (pre-prompt-commit bare conclusions) into a signal that users will read as "ongoing extraction quality." A 57.6% bare-conclusion rate looks like the LLM is broken; in reality the live extraction rate (post-2026-04-20) is ~3% (1 bare fact out of ~30+ created since the prompt commit landed).
+
+The 84% rejection rate has a similar structural issue: it counts cleanup of a bursty `/study-repo` regression against the active-facts denominator, not against the actual extraction quality of the live window.
+
+**Quick fix shipping now (this session):** restrict `quality_score` and the digest's "Quality" section to facts created within the same 30-day window already used by `token_budget`. Surface a separate "historical" line so users can see both numbers, but the headline is the live one. This makes the metric actionable: a high live bare-conclusion rate = live LLM calibration drift; a high historical rate = legacy data, not a current alarm.
+
+**Deferred to 0.12 / 1.x:**
+
+1. The systemic `/study-repo` misattribution failure mode (cause 4) deserves its own guard. External-project READMEs being studied should land in `reference` predicates, not as `uses_database`/`deployment_platform`. Track this as a follow-up entry.
+2. A backfill/cleanup pass on the 34 historical bare-conclusion facts: either retroactive rejection, or a one-shot reclassification that moves them to a `legacy_observation` predicate that the prompt's reason-clause requirement doesn't apply to.
+3. The metric's calibration assumes "bare conclusion = bad", but spot-checking shows several flagged facts are perfectly informative ("MCP tools return dual content + structuredContent via TextSummary module") — they describe mechanics implicitly. The vocabulary may itself be too strict; revisit during 1.0 soak with real usage data.
+
+**Process win:** the metric did its job — it surfaced a real signal that would otherwise have stayed invisible, and the investigation distinguished historical noise from live calibration. Without #48 we'd have no way to know.
 
 ---
 
 ## Executive Summary
 
-The codebase has grown from 11,392 to 12,239 LOC since the Mar 9 review. A configurable embedding provider system was added (+847 LOC across 5 new files and 8 modified files). The three watch-list files have grown further: `tools.rb` (728→745), `recall.rb` (681→727), `sqlite_store.rb` (547→547, unchanged). All three remain above 500 lines.
+Six days, +2,011 LOC. The headline finding: **the watch-list item from 2026-04-22 (#28 — extract per-endpoint helpers from `Dashboard::API`) was not just deferred, it actively regressed.** `dashboard/api.rb` grew from 627 → 807 LOC (+180, +29%), is now the only file in `lib/` over 750 lines, and gained four new methods all exceeding 15 lines. Method-size pressure increased: the previous worst case (`recall` at 39 lines) is now `timeline` at 52 lines, and the file has 11 methods over 15 lines (vs 11 last review) but with a higher mean.
+
+The codebase is otherwise healthy. Five **new dashboard subsystems** (`moments.rb`, `reuse.rb`, `trust.rb`, `scoped_fact_resolver.rb`, plus `efficacy.rb` carried over) shipped with **direct spec files**. Three new schema migrations (v15/v16/v17) all wrap DDL with idempotent `create_table?` / `add_column` and have **per-migration specs plus round-trip specs from v12, v13, and v14 forward to v17** (a deliberate process improvement noted in `feedback_round_trip_migration_specs.md`). Four new sweep operations (`dedupe_open_conflicts`, `reclassify_references`) gained spec coverage in `sweep/maintenance_spec.rb`.
+
+**What regressed:**
+- `dashboard/api.rb` 627 → 807 LOC (+180). Watch-list item not addressed.
+- `sweep/maintenance.rb` 334 → 456 LOC (+122). Two of the four new methods are 50+ lines (`dedupe_open_conflicts` 58, `restore_multi_value_supersessions` 57).
+- Sleep-based test latency grew. `dashboard/moments_spec.rb` and `dashboard/api_spec.rb` add 4 more `sleep 1.1` calls (+4.4s wall). Total sleep-based test cost in suite is now ~8.4s, up from ~4s.
+- One new code smell: `digest_command.rb:128` calls `Dashboard::Trust.new(manager).send(:utilization)` — reaches into a private method instead of exposing `utilization` on the public Trust API.
 
-Five items were resolved in the Mar 19 quality update session: the Shortcuts scope bug (correctness), ApiAdapter exception typing, silent exception logging (3 locations), dead Configuration accessors removed, and `index_database` decomposed into focused methods. No known correctness bugs remain.
+**What was resolved or improved since 2026-04-22:**
+- Round-trip migration specs from v12/v13/v14 → v17 added (release-blocker per `feedback_round_trip_migration_specs.md`).
+- Per-migration specs for v13–v17 added under `spec/claude_memory/store/migrations/`.
+- New dashboard subsystems shipped *with* specs (good pattern — Reuse, Moments, Trust, Knowledge, ScopedFactResolver all have direct specs).
+- `lib/claude_memory/store/sqlite_store.rb` only grew 40 LOC (544 → 584); regrowth controlled.
 
-**Resolved this session:** 13 items (#1 Tools god object, #2 Recall legacy/dual strategy, #3 Shortcuts scope, #4 SQLiteStore extraction, #9 test coverage for critical files, NEW-1 ApiError, NEW-2 dead accessors, NEW-3 silent rescues, #5 index_database decomposition, #6 promote_fact transaction, #7 provenance nil content_item_id, #12 Resolver mutable state, #19 SnippetExtractor DRY)
-**Resolved since last review:** 3 additional items (ExportCommand N+1, `discover_other_projects` bare rescue, embedding test coverage)
-**Total remaining:** 12 items (0 high, 5 medium, 4 low, 5 carried forward — all structural work complete)
+**New this review:** 4 items. 1 high-priority (Dashboard::API extraction, now urgent), 1 medium (`sweep/maintenance.rb` size), 1 low (`Time.parse` duplication across dashboard files), 1 quick win (`.send(:utilization)` smell in digest).
 
 ### Current Strengths
 
-- Functional core: 20+ pure logic classes with zero I/O
-- Domain objects: properly frozen and self-validating
-- Null object pattern: NullFact, NullExplanation
-- Result monad: Core::Result for Success/Failure
-- 100% frozen_string_literal compliance (117 files)
-- 1.84:1 test-to-code ratio (22,563 spec : 12,239 lib)
-- New embedding subsystem: shared RSpec examples, duck-typed providers, Data.define value object
-- Zero N+1 patterns in hot paths
-- Proper batch loading via FactQueryBuilder
-- Content-addressed dedup in IndexCommand
-- DimensionCheck value object (functional core, no side effects)
-- Zero known correctness bugs
+- Migrations now ship with per-migration specs **and** cross-version round-trip specs — a deliberate release-readiness improvement that landed during this window
+- New dashboard subsystems all have direct specs; spec count grew 156 → 188 files (+32)
+- Domain objects, frozen string literals, transaction wrapping, no raw SQL, no N+1 in hot paths — all preserved
+- Five files >300 LOC last review; eight now, but mostly because of new modules carrying single responsibilities (Moments 244, Trust 284, Conflicts 285), not god-object regrowth
 
 ---
 
 ## 1. Sandi Metz Perspective
 
 ### What's Been Fixed ✅
-- New embedding providers follow duck typing contract (no base class inheritance)
-- Shared RSpec examples verify provider contract (`spec/support/shared_examples/embedding_provider.rb`)
-- `set_meta`/`get_meta` promoted to public API (needed by DimensionCheck, VectorIndex)
-- ExportCommand N+1 eliminated with batch loading
-- **Shortcuts scope bug fixed** — scopes changed from symbols to strings to match DualQueryTemplate comparisons
-- **`index_database` decomposed** — split 130-line method into 5 focused methods: `index_database` (orchestrator), `handle_dimension_mismatch`, `find_facts_to_index`, `run_indexing`, `process_batch`, `report_dedup_stats`
-- **Tools god object eliminated** — split 745-line `Tools` class into thin 104-line dispatcher + 6 handler modules: `QueryHandlers` (90), `ShortcutHandlers` (37), `ContextHandlers` (38), `ManagementHandlers` (124), `StatsHandlers` (188), `SetupHandlers` (211). No file exceeds 211 lines. Public API unchanged.
-- **Recall legacy/dual duplication eliminated** — split 727-line `Recall` into thin 94-line facade + strategy engines: `DualEngine` (101), `LegacyEngine` (134), shared `QueryCore` module (357). The `if @legacy_mode` branching in every public method is replaced by engine delegation. Public API unchanged.
+
+- `SQLiteStore` regrowth held steady at 584 LOC after the 2026-04-22 LLMCache + MetricsAggregator extractions; only +40 LOC over 6 days, and that's adding two new tables (`moment_feedback`, `activity_events`) with their CRUD wrappers
+- New dashboard subsystems each landed under 300 LOC with focused responsibilities
+  - `Moments` (244 LOC) — feed-shape construction, no DB writes
+  - `Trust` (284 LOC) — sidebar aggregations, all reads
+  - `Reuse` (97 LOC) — top-N "most-used" panel
+  - `Knowledge` (136 LOC) — fact summary panel
+  - `ScopedFactResolver` (95 LOC) — pure helper
+- Round-trip migration specs (`round_trip_v12_to_v17_spec.rb` etc.) — Sandi-style "test the contract, not the implementation"
 
 ### Critical Issues 🔴
 
-None remaining.
+#### A. `Dashboard::API` regressed: 627 → 807 LOC (+29%) — **carried-forward item became urgent**
+
+`lib/claude_memory/dashboard/api.rb` was the watch-list item at the close of the 2026-04-22 review (#28). Six days later, instead of shrinking via per-endpoint extraction, it absorbed:
+
+- `find_recall_trigger` (lib/claude_memory/dashboard/api.rb:193) — 32 lines, 5 SQL constructions, calls 3 helpers, JSON-parses event details
+- `extract_user_prompt` (lib/claude_memory/dashboard/api.rb:237) — 29 lines, JSONL parsing, content type narrowing, plumbing-noise filtering
+- `facts` (lib/claude_memory/dashboard/api.rb:373) — 39 lines (was 26), now also handles `stale_only` filtering with cross-store exclusion
+- `facts_seen_in_recent_recalls` (lib/claude_memory/dashboard/api.rb:418) — 20 lines, scoped-pair aggregation
+- `efficacy` (lib/claude_memory/dashboard/api.rb:439) — 31 lines (was 23), now branches on session_id with time-window correlation
+- New micro-endpoints: `moments`, `trust`, `knowledge`, `reuse`, `moment_feedback`, `clear_moment_feedback`, `fact_detail`, `promote_fact`, `reject_fact`
+
+The class now has **42 methods** (up from ~31) and **8 methods over 20 lines**. The methods that delegate cleanly (`conflicts`, `moments`, `trust`, `knowledge`, `reuse` — all 1-liners) are the right pattern; the rest of the file should follow that pattern.
+
+**Method size table (current state):**
+
+| Method | Line | Size | Concern |
+|---|---|---|---|
+| `timeline` | 471 | 52 | 3 separate Sequel aggregations + Ruby-side merge — should be `Dashboard::Timeline` |
+| `vec_health` | 759 | 46 | Branchy status derivation over coverage stats |
+| `recall` | 315 | 41 | Result flattening + bare rescue + actionable-hint branching |
+| `facts` | 373 | 39 | Pagination + filter + cross-store stale exclusion |
+| `activity_detail` | 149 | 37 | Joined fetch + linked facts + recall-trigger correlation |
+| `hooks_health` | 704 | 32 | Multi-state status with fix messages |
+| `find_recall_trigger` | 193 | 32 | Time-window query with session_id fallback |
+| `efficacy` | 439 | 31 | Session-scope vs window-scope branching |
+| `extract_user_prompt` | 237 | 29 | JSONL reverse-walk + plumbing filter |
+| `session_summary` | 119 | 29 | Multi-event-type aggregation |
+| `db_stats` | 647 | 28 | Predicate counts + entity counts + size stats |
+
+**Proposed extractions** (each candidate is testable in isolation):
+
+```ruby
+# lib/claude_memory/dashboard/timeline.rb — pure aggregation
+class Timeline
+  def initialize(manager) = @manager = manager
+  def days = { days: build_days }
+  private
+  def build_days
+    return [] unless store
+    fact_rows, content_rows, event_rows = load_aggregations
+    merge_into_days(fact_rows, content_rows, event_rows)
+  end
+end
+
+# lib/claude_memory/dashboard/health.rb — already 4 health checks (db, hooks, vec, vectors)
+class Health
+  def report = { status: overall(checks), checks: checks, version: VERSION }
+  private
+  def checks = [db_health("global"), db_health("project"), hooks_health, vec_health]
+end
+
+# lib/claude_memory/dashboard/recall_query.rb — wraps live recall + actionable error mapping
+class RecallQuery
+  def call(params) = format_response(run(params))
+end
+
+# lib/claude_memory/dashboard/recall_trigger_finder.rb — pure time-window correlation
+# lib/claude_memory/dashboard/user_prompt_extractor.rb — pure JSONL parsing
+# lib/claude_memory/dashboard/facts_query.rb — pagination + stale exclusion
+```
+
+After these extractions `api.rb` should drop to **~250 LOC** of routing-and-delegation. The pattern was already proven by `Conflicts` / `Moments` / `Trust` / `Knowledge` / `Reuse`.
+
+**File:** `lib/claude_memory/dashboard/api.rb`
+**Effort:** 4–6 hours (5 extractions, each with a focused spec)
+**Priority:** 🔴 — was medium last review, escalates to high because the trend line points at 1,000+ LOC by next sprint if uncorrected
+**Expert principle:** Sandi Metz SRP; Bernhardt boundaries; Beck simple design
 
-### High Priority Issues
+### Medium Issues 🟡
+
+#### B. `sweep/maintenance.rb` grew 334 → 456 LOC (+122, +37%)
+
+Last review noted maintenance.rb at 334 (after dropping from 456 earlier — see the review's appendix B). It's now back at 456. Two large methods landed:
 
-None remaining. All >500-line files have been decomposed.
+- `dedupe_open_conflicts` (lib/claude_memory/sweep/maintenance.rb:273) — 58 lines, multi-step transaction (group → resolve duplicates → reattach provenance → reject losers → mark conflicts resolved)
+- `reclassify_references` (lib/claude_memory/sweep/maintenance.rb:340) — 26 lines, transactional cleanup that requires `Distill::ReferenceMaterialDetector`
+
+Plus the pre-existing `restore_multi_value_supersessions` (line 185, 57 lines).
+
+These are all *one-shot historical cleanups* (per their docstrings). They don't belong in the regular sweep cycle — they're admin operations. Two options:
+
+1. **Extract to `Sweep::HistoricalCleanup`** — a separate module for one-shot data fixes
+2. **Keep in Maintenance but extract long methods** — e.g. `dedupe_open_conflicts` calls `pair_key`, but the inner per-group logic (lib/claude_memory/sweep/maintenance.rb:294-326) is 32 lines that could be `resolve_duplicate_group(keeper, duplicates)`
+
+**File:** `lib/claude_memory/sweep/maintenance.rb`
+**Effort:** 2 hours
+**Priority:** 🟡 Medium
+**Expert principle:** Sandi Metz SRP; Beck single level of abstraction
+
+#### C. `digest_command.rb:128` calls private API via `.send`
+
+```ruby
+# lib/claude_memory/commands/digest_command.rb:128
+util = Dashboard::Trust.new(manager).send(:utilization)
+```
+
+This is the only `.send` to a private method in `lib/`. Two paths forward:
+
+```ruby
+# Option 1: Promote utilization to public on Trust (it already returns a documented Hash shape)
+# lib/claude_memory/dashboard/trust.rb — remove `private` annotation above utilization
+
+# Option 2: Extract Dashboard::Utilization as its own object
+class Utilization
+  def initialize(manager) = @manager = manager
+  def report = { extracted:, used:, used_from_extracted:, ratio_pct:, window_days: }
+end
+```
+
+Option 2 is cleaner — Trust currently *also* exposes `utilization` indirectly through `snapshot`, so users have two paths to the same data. Extracting the calculator gives Digest, Trust, and any future caller one canonical interface.
+
+**File:** `lib/claude_memory/commands/digest_command.rb:128`
+**Effort:** 30 minutes
+**Priority:** 🟡 Medium (works correctly, but tells future readers "private is negotiable")
+**Expert principle:** Avdi Grimm tell-don't-ask; Sandi Metz dependency clarity
+
+### Low Issues
+
+| # | Issue | File:Line | Effort |
+|---|---|---|---|
+| 8 | `upsert_content_item` 11 keyword params (carried) | `store/sqlite_store.rb:193` | 1 hour |
+| 32 | `parse_timestamp` duplicated in `dashboard/api.rb:565` and `dashboard/conflicts.rb:278` | both | 15 min |
+| 33 | `stores_for(scope)` / `facts_stores_for(scope)` near-identical pattern | `dashboard/conflicts.rb:160`, `dashboard/api.rb:589` | 30 min |
 
 ---
 
 ## 2. Jeremy Evans Perspective
 
 ### What's Been Fixed ✅
-- Batch queries in Recall pipeline (FactQueryBuilder)
-- Transaction wrapping in Resolver
-- ExportCommand N+1 eliminated
-- `discover_other_projects` now catches specific exception types (`Sequel::DatabaseError, Extralite::Error, IOError`)
-- **promote_fact transaction boundary fixed** — project data read before global transaction (already correct in code; verified and confirmed)
-- **Provenance nil content_item_id fixed** — removed mandatory `content_item_id` validation from `Domain::Provenance`, allowing nil for promoted facts
 
-### Medium Issues 🟡
+- Migrations v15, v16, v17 all wrap DDL in idempotent `create_table?` / `add_column` and provide `down` blocks (v14's down is intentionally a no-op with comment)
+- `Trust#extracted_fact_pairs` (lib/claude_memory/dashboard/trust.rb:231) and `used_fact_pairs` (line 248) batch via `select(:id)` + iteration — no per-row queries
+- `Conflicts#load_facts_for_rows` (lib/claude_memory/dashboard/conflicts.rb:235) batches with `where(id: ids).as_hash(:id)` — explicit N+1 prevention
 
-| # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 8 | **upsert_content_item has 11 keyword parameters** | `store/sqlite_store.rb:158-184` | 1 hour |
+### Raw SQL Audit
+
+No new raw SQL. The handful of `Sequel.lit` calls in `dashboard/api.rb` are all `DATE(...)` group-by helpers (lines 479, 487, 494) — required because Sequel doesn't have a portable `DATE(timestamp_string)` extractor for SQLite.
 
-Exceeds the 5-parameter guideline. Suggests the method is doing too much.
+### Transaction Safety
 
-**Fix:** Introduce a `ContentItemAttributes` value object.
+New transactional methods all wrap correctly:
+- `Sweep::Maintenance#dedupe_open_conflicts` — wraps in `@store.db.transaction` (line 289)
+- `Sweep::Maintenance#reclassify_references` — wraps in `@store.db.transaction` (line 349)
+- `SQLiteStore#upsert_moment_feedback` — wraps in `@db.transaction` (line 128)
+
+### N+1 Audit (new dashboard panels)
+
+- `Moments#build_moment` (lib/claude_memory/dashboard/moments.rb:125) calls `resolve_content` and `extracted_facts` per row. **Potential N+1 if a feed page surfaces 50 ingest moments.** `extracted_facts` runs `store.db[:facts].join(:provenance).where(content_item_id:)` per moment.
+- `Trust#count_open_conflicts` (lib/claude_memory/dashboard/trust.rb:145) → `Conflicts#distinct_open_counts` walks both stores. Acceptable (fixed cardinality of 2).
+- `Trust#used_fact_pairs` (lib/claude_memory/dashboard/trust.rb:248) loads up to N=500 events without limit. Could grow unbounded. Recommend explicit `.limit(...)` for safety.
+
+**Recommendation:**
+- Batch `extracted_facts` in `Moments`: collect all `content_item_id`s up front, run one `where(content_item_id: ids)` join, group results in Ruby.
+- Add explicit `.limit` to `used_fact_pairs` (10,000 is a safe ceiling for a 30-day window).
+
+**File:** `lib/claude_memory/dashboard/moments.rb:125,231`
+**Effort:** 45 minutes
+**Priority:** 🟡 Medium (will only bite at scale; fix proactively)
+**Expert principle:** Jeremy Evans dataset hygiene
 
 ---
 
 ## 3. Kent Beck Perspective
 
 ### What's Been Fixed ✅
-- New embedding subsystem has full test coverage (4 spec files, shared examples)
-- `generator_spec.rb` now tests `name`/`dimensions` contract
-- DimensionCheck tested for all 3 states (`:fresh`, `:match`, `:mismatch`)
-- ApiAdapter tested with HTTP mocks (no WebMock dependency)
-- **5 critical untested files now have specs** — `similarity.rb` (10 tests), `metadata_extractor.rb` (9 tests), `tool_extractor.rb` (7 tests), `recover_command.rb` (3 tests), `schema_validator.rb` (6 tests). Total: +36 specs, suite now at 1444.
+
+- **Migration spec coverage hit gold standard.** Per-migration specs for v13/v14/v15/v16/v17 + cross-version round-trips from v12, v13, and v14 all forward to v17. That's the canonical "test the seam" pattern. The lessons from `feedback_round_trip_migration_specs.md` are now codified in green tests.
+- New commands `digest_command.rb` and `census_command.rb` shipped with direct specs
+- New dashboard modules all have direct specs (`moments_spec.rb`, `reuse_spec.rb`, `trust_spec.rb`, `knowledge_spec.rb`, `scoped_fact_resolver_spec.rb`)
 
 ### High Priority Issues
 
-None remaining. All high-priority items are resolved.
+#### D. Two new commands shipped without specs
 
-### Remaining Untested Files (lower priority)
+| Command | LOC | Spec? |
+|---|---|---|
+| `commands/dedupe_conflicts_command.rb` | 55 | ❌ none |
+| `commands/reclassify_references_command.rb` | 56 | ❌ none |
 
-Thin CLI wrappers that delegate to already-tested classes:
-- `commands/stats_command.rb`, `commands/export_command.rb`
-- `commands/changes_command.rb`, `commands/conflicts_command.rb`, `commands/explain_command.rb`
-- `commands/recall_command.rb`, `commands/search_command.rb`
-- `commands/sweep_command.rb`, `commands/publish_command.rb`, `commands/ingest_command.rb`
-- `commands/db_init_command.rb`
-- `commands/checks/` (6 files), `commands/initializers/` (5 files)
-- `mcp/tool_helpers.rb`, `embeddings/fastembed_adapter.rb`, `distill/distiller.rb`
+Both are thin wrappers over `Sweep::Maintenance` (which *is* tested), but the CLI-layer concerns — option parsing, scope routing, output format, dry-run flag flow-through — are uncovered.
 
-### Medium Issues 🟡
+The output format in particular has logic worth pinning:
+- `dedupe_conflicts_command.rb:38-52` decides `DRY RUN` vs `DEDUPE`, separator length, decisions header
+- `reclassify_references_command.rb:38-53` truncates objects to 100 chars + ellipsis
 
-| # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 10 | **Sleep-based tests add 4+ seconds** | `spec/ingest/ingester_spec.rb:43,65,81` | 1 hour |
+**Proposed:** Mirror `digest_command_spec.rb` (or `census_command_spec.rb`) — test option parsing, dry-run paths, and stdout shape via injected `StringIO`.
 
-Three `sleep 1.01` calls wait for filesystem mtime changes. `publish_spec.rb:189` has `sleep 1.1`.
+**Effort:** 30 min each (60 min total)
+**Priority:** High — these are admin commands that mutate data; CLI ergonomics belong under test
 
-**Fix:** Mock `File.mtime` or inject a time provider instead of real sleeps.
+#### E. `dashboard/server.rb` still untested
 
-| # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 11 | **No shared test factory** | `spec/spec_helper.rb` | 1 hour |
+Carried over from 2026-04-22. The file has grown 189 → 211 LOC (+22) due to new endpoints (moments feedback POST/DELETE, conflict reject_similar). All branching is inside the request router (`handle_moments`, `handle_conflicts`).
+
+WEBrick HTTP testing is awkward but not impossible — `Rack::MockRequest` works against the API class directly. Alternatively, exercise the routing by injecting a stub WEBrick request object.
+
+**Effort:** 1.5 hours
+**Priority:** Medium-Low
+
+### Sleep-Based Test Latency Increased
+
+Total sleep-based test cost in `bundle exec rspec`:
+
+| Spec | sleep total | Notes |
+|---|---|---|
+| `spec/claude_memory/ingest/ingester_spec.rb` | 3.03s | mtime resolution, carried |
+| `spec/claude_memory/publish_spec.rb` | 1.1s | carried |
+| `spec/claude_memory/recall_spec.rb` | 0.01s | carried |
+| `spec/claude_memory/dashboard/moments_spec.rb` | 2.2s | **NEW** ordering of activity events |
+| `spec/claude_memory/dashboard/api_spec.rb` | 2.2s | **NEW** activity ordering tests |
+| **Total** | **~8.5s** | up from ~4s last review |
+
+The dashboard sleeps are because activity_events ordering depends on `occurred_at` ISO timestamps, and successive inserts in <1s produce the same timestamp. Two fixes:
+
+```ruby
+# Option 1: Inject explicit timestamps (already supported via insert column)
+store.activity_events.insert(occurred_at: Time.now.utc.iso8601, ...)
+store.activity_events.insert(occurred_at: (Time.now + 1).utc.iso8601, ...)
+
+# Option 2: Stub Time.now via Timecop or RSpec's allow(Time).to receive(:now)
+```
 
-`spec_helper.rb` is only 21 lines. ~20 test files independently define `create_fact` and `create_content_with_fact` helpers. The canonical pattern from `tools_spec.rb:275` should be extracted.
+Option 1 requires no extra dep. Either eliminates 4.4s of wall time.
 
-**Fix:** Create `spec/support/database_factory.rb` with shared helpers, require from spec_helper.
+**File:** `spec/claude_memory/dashboard/moments_spec.rb:130,132`, `api_spec.rb:332,359`
+**Effort:** 30 minutes
+**Priority:** 🟡 Medium (test speed degrades CI loop)
+**Expert principle:** Kent Beck fast feedback
+
+### Carried-Forward Issues 🟡
+
+| # | Issue | File:Line | Effort |
+|---|---|---|---|
+| 11 | No shared test factory | `spec/spec_helper.rb` | 1 hour |
 
 ---
 
 ## 4. Avdi Grimm Perspective
 
 ### What's Been Fixed ✅
-- DimensionCheck returns a Result value object — no exceptions, no side effects
-- `Embeddings.resolve` raises `ArgumentError` with clear message for unknown providers
-- ApiAdapter raises with descriptive messages for missing API keys
-- Duck typing for embedding providers (no base class)
-- **ApiAdapter now uses typed `ApiError < StandardError`** instead of bare `raise "message"`
-- **Resolver mutable state resolved** — verified that `project_path` and `scope` are already threaded as parameters through the entire method chain; no mutable instance state exists
 
-### Carried Forward Issues 🟡
+- New code uses scoped rescues (`rescue Sequel::DatabaseError, JSON::ParserError`) over bare rescues by default. Of 18 new rescue clauses in dashboard files, **13 are scoped to specific exception types**, 5 are bare and all return safe defaults
+- `Result` pattern preserved in embeddings paths
+- `Core::RelativeTime.format` used consistently across new dashboard modules
+
+### Bare Rescue Audit (full lib/, current count: 19 bare rescues)
+
+The count grew from 5 → 19 because new dashboard code added 5 in `api.rb`. All are defensive (return safe shape):
+
+| Location | Context | Returns | Verdict |
+|---|---|---|---|
+| `mcp/handlers/stats_handlers.rb:102` | `fts_legacy?` | `false` | Acceptable — boolean check |
+| `mcp/instructions_builder.rb:147` | `vec_available?` | `false` | Acceptable |
+| `sweep/maintenance.rb:140` | FTS prune | skips row | Acceptable |
+| `commands/hook_command.rb:102` | forked handler | `nil` | Required |
+| `commands/stats_command.rb:276` | `check_fts_format` | no-op | Acceptable |
+| **`dashboard/api.rb:340` (new)** | recall live query | error hash | Acceptable — wide net for unfamiliar errors from Recall pipeline |
+| **`dashboard/api.rb:672` (new)** | `db_stats` aggregation | `{exists:, error:}` | Acceptable |
+| **`dashboard/api.rb:693` (new)** | `db_health` introspection | error hash | Acceptable |
+| **`dashboard/api.rb:728` (new)** | `hooks_health` JSON read | error hash | Acceptable |
+| **`dashboard/api.rb:797` (new)** | `vec_health` | error hash | Acceptable |
+
+Verdict: per `Style/RescueStandardError` in Standard Ruby (rejected explicit-rescue change in last review), these are correct. **No action.**
+
+### Carried-Forward Issues 🟡
 
 | # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 13 | **Inconsistent payload validation in hooks** | `hook/handler.rb:17-53` | 30 min |
+|---|---|---|---|
+| 13 | Inconsistent payload validation | `hook/handler.rb:53-82` | 30 min |
+
+Verified still present.
+
+### New Concern
+
+#### F. `digest_command.rb:128` reaches into `Trust`'s private API
 
-`ingest` uses `.fetch("field")` with fallback, `sweep` uses `.fetch("budget", default)`, `publish` uses `.fetch("mode", "shared")`. No consistent validation pattern.
+Documented above (#C). Repeating here under the Avdi lens: the explicit `.send` is a public-API smell. Either the method shouldn't be private, or there should be a public wrapper. Choose.
 
 ---
 
 ## 5. Gary Bernhardt Perspective
 
 ### What's Been Fixed ✅
-- DimensionCheck is pure: takes store + provider, returns immutable Result. No hidden side effects.
-- `clear_stale_embeddings` was moved from hidden infrastructure setup to explicit command-level call.
-- VectorIndex#clear! encapsulates vec0 table knowledge (no raw SQL in command).
-- **Dead Configuration embedding accessors removed** — resolver and ApiAdapter read ENV directly, no unused indirection.
 
-### Carried Forward Issues 🟡
+- New dashboard modules continue to honor the imperative-shell / functional-core split:
+  - `Trust` does only reads + transformation (no writes)
+  - `Moments` does reads + transformation
+  - `Reuse` does reads + transformation
+  - `Efficacy::Reporter` is **pure** (no DB) — takes events, returns a hash — Bernhardt's dream
+- `Knowledge#summary` returns shaped data; UI logic stays out of the model
+- New value-object-y data: `KIND_TO_EVENT_TYPES`, `FEED_EVENT_TYPES` are frozen module constants
 
-| # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 14 | **I/O mixed with logic in discover_other_projects** | `mcp/tools.rb:565-614` | 1 hour |
+### Boundaries
 
-SQL queries, filesystem checks, database connections in a loop, and error handling all mixed together.
+```
+HTTP layer:    Dashboard::Server (211 LOC, untested)         ← imperative shell
+JSON layer:    Dashboard::API (807 LOC ⚠ growing)            ← needs to shrink to routing
+Subsystems:    Conflicts, Moments, Trust, Knowledge, Reuse   ← functional core (good)
+Pure helpers:  Efficacy::Reporter, ScopedFactResolver        ← pure (excellent)
+Query layer:   Recall, store datasets                        ← impure but isolated
+```
 
-| # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 15 | **Sweeper mutable state** | `sweep/sweeper.rb:16-17` | 20 min |
+`API` is the wrong layer to be doing JSONL parsing (`extract_user_prompt`), time-window correlation (`find_recall_trigger`), or 3-source aggregation (`timeline`). Each of those wants to be its own pure object.
+
+### Test Speed Regression
+
+Sleep-based tests are dollar-bills the suite is burning every CI run. Eliminating them is functional-core hygiene — the test should pin behavior, not wait for clock state.
+
+### Carried-Forward Issues 🟡
 
 | # | Issue | File:Line | Effort |
-|---|-------|-----------|--------|
-| 16 | **Dir.chdir in publish tests** | `spec/publish_spec.rb:14` | 15 min |
+|---|---|---|---|
+| 15 | Sweeper mutable state | `sweep/sweeper.rb:16-17` | 20 min |
+| 16 | `Dir.chdir` in publish tests | `spec/publish_spec.rb:14` | 15 min |
 
 ---
 
 ## 6. General Ruby Idioms
 
-### What's Been Fixed ✅
-- **Silent exception swallowing resolved** — 3 bare rescue blocks now log via `ClaudeMemory.logger.debug(...)`:
-  - `mcp/instructions_builder.rb:29`
-  - `hook/context_injector.rb:47`
-  - `commands/checks/vec_check.rb:55`
+### New Items
 
-- **SnippetExtractor range calculation DRY** — extracted `snippet_range` method to eliminate duplicated start/end index computation between `extract_with_lines` and `build_snippet`
-
-### Carried Forward Issues
+| # | Issue | File:Line | Severity | Effort |
+|---|---|---|---|---|
+| 31 | `Dashboard::API` 807 LOC, 11 methods >15 lines (regression of #28) | `dashboard/api.rb` | 🔴 High | 4–6 hours |
+| 32 | `parse_timestamp(value)` duplicated verbatim in api.rb:565 and conflicts.rb:278 | both | 🟢 Low | 15 min |
+| 33 | `stores_for` / `facts_stores_for` near-identical between Conflicts and API | `conflicts.rb:160`, `api.rb:589` | 🟢 Low | 30 min |
+| 34 | `digest_command.rb:128` uses `.send(:utilization)` to call private | `digest_command.rb:128` | 🟡 Medium | 30 min |
+| 35 | Sleep-based dashboard tests add 4.4s to suite | `dashboard/{moments,api}_spec.rb` | 🟡 Medium | 30 min |
+| 36 | DedupeConflictsCommand and ReclassifyReferencesCommand untested | `commands/` | High | 60 min |
+| 37 | `sweep/maintenance.rb` regrew to 456 LOC; 3 methods >50 lines | `sweep/maintenance.rb` | 🟡 Medium | 2 hours |
+| 38 | `Moments#extracted_facts` per-moment join (potential N+1 at 50-row pages) | `moments.rb:231` | 🟡 Medium | 30 min |
+
+### Carried-Forward Items
 
 | # | Issue | File:Line | Severity | Effort |
-|---|-------|-----------|----------|--------|
-| 17 | **ResponseFormatter duplication** | `mcp/response_formatter.rb:27-280` | 🟡 Medium | 1 hour |
-| 18 | **Publish section generator repetition** | `publish.rb:100-154` | Low | 30 min |
+|---|---|---|---|---|
+| 17 | ResponseFormatter duplication | `mcp/response_formatter.rb` | 🟡 Medium | 1 hour |
+| 28 | ~~Dashboard::API method extraction~~ — **escalated to #31** | — | — | — |
+| 8 | `upsert_content_item` 11 keyword params | `store/sqlite_store.rb:193` | 🟢 Low | 1 hour |
+| 10 | Sleep-based ingester tests | `spec/ingest/ingester_spec.rb` | 🟢 Low | 1 hour |
+| 11 | No shared test factory | `spec/spec_helper.rb` | 🟢 Low | 1 hour |
 
 ---
 
 ## 7. Positive Observations
 
-- **Batch loading architecture**: `FactQueryBuilder` and `BatchLoader` eliminate N+1 patterns in all hot query paths
-- **Consistent dependency injection**: All commands accept `stdout`, `stderr`, `stdin` for testability
-- **Clean module boundaries**: Each module has clear responsibilities with minimal cross-coupling
-- **Proper Sequel usage**: Datasets used consistently, raw SQL avoided almost entirely
-- **Excellent domain modeling**: Fact, Entity, Provenance are immutable value objects with validation
-- **Good file organization**: ~1 class per file, consistent naming, clear module nesting
-- **Strong test culture**: 1.84:1 test-to-code ratio, behavior-focused tests
-- **Infrastructure abstractions**: `FileSystem`, `InMemoryFileSystem` enable fast tests
-- **Core::Result monad**: Consistent Success/Failure pattern throughout
-- **New embedding subsystem**: Clean duck typing with shared RSpec examples verifying provider contract. DimensionCheck is a textbook value object — pure function, immutable result, no side effects. The resolver uses simple case/when (no over-engineered factory/registry).
-- **VectorIndex#clear!**: Properly encapsulates destructive vec0 operation behind the abstraction boundary
+- **Migration discipline** — round-trip specs, per-migration specs, idempotent DDL. The "treat round-trip migration specs as a release blocker" lesson from `feedback_round_trip_migration_specs.md` got operationalized in 5 days
+- **New commands ship with specs** — DigestCommand and CensusCommand both got direct specs; the two that didn't (Dedupe + Reclassify) are 55-line wrappers over already-tested Maintenance methods, so the gap is small
+- **Dashboard subsystem decomposition** — when 5 new panels (Moments, Reuse, Trust, Knowledge, ScopedFactResolver) all land as their own classes with their own specs, the module-extraction muscle is strong
+- **`Efficacy::Reporter` purity** — 128 LOC, zero I/O, takes events and returns shape. Spec is fast and readable. This is the model the rest of dashboard/ should converge on
+- **No raw SQL added; no N+1 in hot paths; transaction safety maintained** — across +2,011 LOC in 6 days
 
 ---
 
 ## 8. Priority Refactoring Recommendations
 
-### High Priority (Next Week)
+### High Priority (This Week — pre-0.10.0 release)
 
-None remaining. All high-priority items are resolved.
+| # | Item | File:Line | Effort | Impact |
+|---|---|---|---|---|
+| 31 | Extract `Dashboard::Timeline` / `Health` / `RecallQuery` / `RecallTriggerFinder` / `UserPromptExtractor` / `FactsQuery` from API | `dashboard/api.rb` | 4–6 hours | API drops 807→~250 LOC; reverses regression |
+| 36 | Add `dedupe_conflicts_command_spec.rb` + `reclassify_references_command_spec.rb` | `spec/claude_memory/commands/` | 1 hour | CLI surface tested |
 
 ### Medium Priority (Next Sprint)
-| # | Item | Effort | Impact |
-|---|------|--------|--------|
-| 8 | ContentItemAttributes value object | 1 hour | Readability |
-| 10 | Replace sleep-based tests with mocks | 1 hour | Test speed |
-| 11 | Shared test factory | 1 hour | DRY |
-| 17 | ResponseFormatter base method | 1 hour | DRY |
-| 14 | Separate I/O in discover_other_projects | 1 hour | Boundaries |
+
+| # | Item | File:Line | Effort |
+|---|---|---|---|
+| 34 | Promote `Trust#utilization` to public OR extract `Dashboard::Utilization` | `dashboard/trust.rb`, `digest_command.rb:128` | 30 min |
+| 35 | Replace sleep-based dashboard tests with explicit timestamps | `dashboard/{moments,api}_spec.rb` | 30 min |
+| 37 | Extract long methods from `sweep/maintenance.rb` (`dedupe_open_conflicts`, `restore_multi_value_supersessions`) OR move one-shot cleanups to `Sweep::HistoricalCleanup` | `sweep/maintenance.rb` | 2 hours |
+| 38 | Batch `Moments#extracted_facts` to avoid 50-row N+1 | `moments.rb:231` | 30 min |
+| 17 | ResponseFormatter consolidation (carried) | `mcp/response_formatter.rb` | 1 hour |
+| 13 | Payload validator for hook events (carried) | `hook/handler.rb` | 30 min |
+| E | `dashboard/server_spec.rb` (carried) | `spec/claude_memory/dashboard/` | 1.5 hours |
 
 ### Low Priority (Later)
-| # | Item | Effort | Impact |
-|---|------|--------|--------|
-| 13 | Payload validator for hooks | 30 min | Consistency |
-| 15 | Sweeper mutable state | 20 min | Immutability |
-| 16 | Dir.chdir in tests | 15 min | Test isolation |
-| 18 | Publish section builder | 30 min | DRY |
-
-### Carried Forward (Low Priority from Earlier Reviews)
-| # | Item | Original # |
-|---|------|-----------|
-| 20 | DateTime migration (string timestamps) | Feb 4 #17 |
-| 21 | Command manager helper (`with_manager`) | Feb 4 #19 |
-| 22 | release_connections polymorphism | Feb 4 #20 |
-| 23 | Provenance batch insert (`multi_insert`) | Feb 4 #22 |
-| 25 | Result objects for all queries | Feb 4 #24 |
+
+| # | Item | Effort |
+|---|---|---|
+| 32 | DRY `parse_timestamp` (`api.rb:565` ↔ `conflicts.rb:278`) | 15 min |
+| 33 | DRY `stores_for` / `facts_stores_for` | 30 min |
+| 8 | `ContentItemAttributes` value object | 1 hour |
+| 10 | Replace sleep-based ingester tests | 1 hour |
+| 11 | Shared test factory | 1 hour |
+| 15 | Sweeper mutable state | 20 min |
+| 16 | `Dir.chdir` in publish tests | 15 min |
+
+### Quick Wins (Today)
+
+| # | Item | Effort |
+|---|---|---|
+| 32 | Extract `parse_timestamp` to `Core::RelativeTime` (it already lives there as a value module) | 15 min |
+| 34 | Promote `Trust#utilization` to public | 5 min |
+| 35 | Inject timestamps into dashboard spec inserts | 30 min |
 
 ---
 
 ## 9. Conclusion
 
-The codebase maintains its strong architectural foundation. Nine quality items were resolved this session across two passes: the Shortcuts scope correctness bug, ApiAdapter exception typing, silent exception logging, dead code removal, method decomposition, promote_fact transaction verification, provenance nil validation fix, Resolver state verification, and SnippetExtractor DRY extraction.
+In 6 days the codebase grew 12% (+2,011 LOC). Most of that growth was healthy — five new dashboard subsystems with specs, three migrations with both per-version and round-trip specs, two new admin commands wrapping already-tested maintenance methods. Migration discipline in particular leveled up: the lesson from `feedback_round_trip_migration_specs.md` shipped as actual release-blocking spec coverage.
+
+**The headline regression is `Dashboard::API`.** Last review marked it medium-priority for per-endpoint extraction. Six days later it's gained 180 LOC, four new methods over 15 lines, and one method (`timeline`) that's now 52 lines. This is the file that most rewards extraction — it's already surrounded by collaborators (`Conflicts`, `Moments`, `Trust`, `Knowledge`, `Reuse`) that prove the per-endpoint pattern works. Doing the extraction now reverses the trend; deferring lets it accumulate another 200 LOC by next review.
+
+**Recommended next-action set, in order:**
 
-No known correctness bugs remain. All three >500-line god objects have been eliminated: `Tools` (745→104), `Recall` (727→94), `SQLiteStore` (547→386). Zero critical or high-priority issues remain. All critical untested files now have specs (+36 tests). The remaining 12 items are medium/low priority (thin CLI wrappers, DRY improvements, carried-forward items).
+1. **`/quality-update`** to apply #31 (Dashboard::API extraction) and #36 (missing command specs). Target: api.rb ≤ 300 LOC, all commands tested.
+2. Quick wins #32 + #34 + #35 in the same session (~75 min total).
+3. Schedule #37 and #38 for the next sprint — neither is urgent but both compound if left alone.
+4. After #31 lands, `/review-for-quality` again pre-0.10.0 release to confirm the regression closed.
+
+The 0.10.0 release should not ship with `dashboard/api.rb` at 807 LOC — the per-endpoint extraction is well-defined, well-precedented, and small-batch (5 extractions × ~1hr each). Doing it before tag is the difference between landing 0.10.0 with a healthy dashboard subsystem vs. burying tech debt in the headline feature of the release.
 
 ---
 
 ## Appendix A: Metrics Comparison
 
-| Metric | Jan 29 | Feb 4 | Mar 9 | Mar 19 |
-|--------|--------|-------|-------|--------|
-| Ruby files (lib) | ~85 | 104 | 112 | **117** |
-| LOC (lib) | ~8,000 | 9,982 | 11,392 | **12,239** |
-| LOC (spec) | — | 17,693 | 21,632 | **22,563** |
-| Pure logic classes | 17+ | 20+ | 20+ | **22+** |
-| Test files | 74+ | 98 | 128 | **122** |
-| Test-to-code ratio | ~1.5:1 | 1.77:1 | 1.90:1 | **1.84:1** |
-| Files >500 lines | 0 | 2 | 3 | **0** ✅ |
-| Bare rescues (silent) | 0 | 0 | 1 | **0** ✅ |
-| N+1 patterns (hot paths) | 0 | 0 | 0 | **0** ✅ |
-| N+1 patterns (cold paths) | — | — | 1 | **0** ✅ |
-| Untested lib files | — | — | 16 | **~7 critical** ✅ |
-| Known correctness bugs | — | — | — | **0** ✅ |
+| Metric | Mar 9 | Mar 19 | Apr 22 (review) | Apr 22 (after update) | **Apr 28 (this review)** |
+|---|---|---|---|---|---|
+| Ruby files (lib) | 112 | 117 | 148 | 150 | **161** (+11 new modules) |
+| LOC (lib) | 11,392 | 12,239 | 17,014 | 17,031 | **19,025** (+2,011) |
+| LOC (spec) | 21,632 | 22,563 | 28,074 | 28,490 | **31,079** (+2,605) |
+| Spec files | 128 | 122 | 154 | 156 | **188** (+32) |
+| Test-to-code ratio | 1.90:1 | 1.84:1 | 1.65:1 | 1.67:1 | **1.63:1** ⬇️ |
+| Files >500 lines | 3 | 0 | 2 | 1 | **2** ⬆️ (api.rb 807, sqlite_store.rb 584) |
+| Files >300 lines | 9 | 9 | 10 | 8 | **8** (same count, different mix) |
+| Bare rescues (justified) | 1 | 0 | 5 | 5 | **19** (14 new, all defensive) |
+| Bare rescues (unsafe) | 0 | 0 | 0 | 0 | **0** ✅ |
+| N+1 patterns (hot paths) | 0 | 0 | 0 | 0 | **0** ✅ |
+| Pure logic classes | 20+ | 22+ | 25+ | 27+ | **32+** (+5 new dashboard modules) |
+| Migration round-trip specs | 0 | 0 | 0 | 0 | **3** (v12→v17, v13→v17, v14→v17) ✅ |
+| Per-migration specs | 0 | 0 | 0 | 0 | **13** (001–017 minus a few) ✅ |
+| Sleep-based test cost | — | — | ~4s | ~4s | **~8.5s** ⬆️ |
+| Untested new commands | — | — | 0 | 0 | **2** (dedupe-conflicts, reclassify-references) |
+| Known correctness bugs | — | 0 | 0 | 0 | **0** ✅ |
 
 ## Appendix B: File Size Report
 
-| File | Mar 9 | Mar 19 | Trend |
-|------|-------|--------|-------|
-| `mcp/tools.rb` | 728 | **104** | ⬇️ -624 (extracted to 6 handler modules) |
-| `recall.rb` | 681 | **94** | ⬇️ -587 (extracted to engine + query_core) |
-| `store/sqlite_store.rb` | 547 | **386** | ⬇️ -161 (extracted retry_handler + schema_manager) |
-| `mcp/response_formatter.rb` | 394 | 396 | ⬆️ +2 |
-| `mcp/tool_definitions.rb` | 303 | 334 | ⬆️ +31 |
-| `commands/index_command.rb` | 224 | 272 | ⬆️ +48 |
-| `mcp/text_summary.rb` | 257 | 258 | ⬆️ +1 |
-| `commands/stats_command.rb` | 239 | 250 | ⬆️ +11 |
-| `commands/uninstall_command.rb` | 226 | 226 | — |
-| `publish.rb` | 221 | 221 | — |
-| `infrastructure/schema_validator.rb` | 215 | 215 | — |
-| `commands/hook_command.rb` | 214 | 214 | — |
-| `resolve/resolver.rb` | — | 195 | new to watch |
-| `index/vector_index.rb` | — | 184 | new to watch |
+| File | Mar 19 | Apr 22 (review) | Apr 22 (after update) | **Apr 28 (this review)** | Trend |
+|---|---|---|---|---|---|
+| `dashboard/api.rb` | — | 627 🆕 | 627 | **807** | ⬆️ +180 (+29%) — **regression** |
+| `store/sqlite_store.rb` | 386 | 683 | 544 | **584** | ⬆️ +40 (new tables) |
+| `mcp/tool_definitions.rb` | 334 | 459 | 459 | **459** | — |
+| `sweep/maintenance.rb` | — | 334 | 334 | **456** | ⬆️ +122 — new |
+| `mcp/response_formatter.rb` | 396 | 397 | 397 | **397** | — |
+| `commands/stats_command.rb` | 250 | 346 | 346 | **383** | ⬆️ +37 |
+| `recall/query_core.rb` | 357 | 371 | 371 | **371** | — |
+| `mcp/text_summary.rb` | 258 | 313 | 313 | **313** | — |
+| `dashboard/conflicts.rb` | — | 195 | 195 | **285** | ⬆️ +90 (dedup grouping logic) |
+| `dashboard/trust.rb` | — | — | — | **284** | 🆕 new feed-first sidebar |
+| `resolve/resolver.rb` | 195 | 254 | 254 | **268** | ⬆️ +14 (dedupe + scope_hint fix) |
+| `mcp/tools.rb` | 104 | 249 | 249 | **264** | ⬆️ +15 |
+| `commands/index_command.rb` | 272 | 259 | 259 | **259** | — |
+| `commands/hook_command.rb` | 214 | 215 | 215 | **249** | ⬆️ +34 |
+| `publish.rb` | 221 | 256 | 248 | **248** | — |
+| `dashboard/moments.rb` | — | — | — | **244** | 🆕 feed primitive |
+| `commands/uninstall_command.rb` | 226 | 226 | 226 | **226** | — |
+| `hook/context_injector.rb` | — | 214 | 214 | **225** | ⬆️ +11 |
+| `store/store_manager.rb` | — | 215 | 215 | **215** | — |
+| `infrastructure/schema_validator.rb` | 215 | 215 | 215 | **215** | — |
+| `commands/census_command.rb` | — | — | — | **210** | 🆕 predicate census |
+| `mcp/handlers/setup_handlers.rb` | 211 | 211 | 211 | **211** | — |
+| `dashboard/server.rb` | — | 189 | 189 | **211** | ⬆️ +22 (new endpoints) |
+| `embeddings/model_registry.rb` | — | — | — | **210** | 🆕 |
+| `mcp/server.rb` | — | 206 | 206 | **206** | — |
+| `mcp/handlers/stats_handlers.rb` | — | 205 | 205 | **205** | — |
+| `commands/initializers/hooks_configurator.rb` | — | — | — | **200** | — |
+| `commands/embeddings_command.rb` | — | — | — | **198** | — |
+| `ingest/ingester.rb` | — | — | — | **190** | — |
+| `index/vector_index.rb` | 184 | 184 | 184 | **184** | — |
+| `commands/digest_command.rb` | — | — | — | **181** | 🆕 weekly digest |
+| `mcp/handlers/management_handlers.rb` | — | — | — | **177** | — |
+| `ingest/observation_compressor.rb` | — | — | — | **177** | 🆕 tool-specific compression |
+| `recall.rb` | 94 | 175 | 175 | **175** | — |
+| `core/fact_query_builder.rb` | — | — | — | **174** | — |
+| `mcp/error_classifier.rb` | — | — | — | **171** | — |
+| `embeddings/generator.rb` | — | — | — | **165** | — |
+| `index/lexical_fts.rb` | — | — | — | **153** | — |
+| `dashboard/knowledge.rb` | — | — | — | **136** | 🆕 |
+| `dashboard/efficacy.rb` | — | 127 | 127 | **127** | — |
+| `dashboard/fact_presenter.rb` | — | 109 | 109 | **109** | — |
+| `dashboard/reuse.rb` | — | — | — | **97** | 🆕 |
+| `dashboard/scoped_fact_resolver.rb` | — | — | — | **95** | 🆕 |
+| `commands/reclassify_references_command.rb` | — | — | — | **56** | 🆕 (untested) |
+| `commands/dedupe_conflicts_command.rb` | — | — | — | **55** | 🆕 (untested) |
+
+## Appendix C: Methods >15 Lines in Watch-List Files
+
+### `dashboard/api.rb` (807 LOC, **42 methods**)
+
+| Method | Line | Size | Action |
+|---|---|---|---|
+| `timeline` | 471 | 52 | Extract `Dashboard::Timeline` |
+| `vec_health` | 759 | 46 | Extract into `Dashboard::Health` |
+| `recall` | 315 | 41 | Extract `Dashboard::RecallQuery` |
+| `facts` | 373 | 39 | Extract `Dashboard::FactsQuery` |
+| `activity_detail` | 149 | 37 | Extract event-detail builder |
+| `hooks_health` | 704 | 32 | Extract into `Dashboard::Health` |
+| `find_recall_trigger` | 193 | 32 | Extract `Dashboard::RecallTriggerFinder` |
+| `efficacy` | 439 | 31 | Move session-window logic into `Efficacy::Loader` |
+| `extract_user_prompt` | 237 | 29 | Extract `Dashboard::UserPromptExtractor` |
+| `session_summary` | 119 | 29 | Extract aggregator |
+| `db_stats` | 647 | 28 | Extract into `Dashboard::Health` |
+| `db_health` | 676 | 25 | Extract into `Dashboard::Health` |
+| `load_content_item` | 603 | 21 | Could move into `FactPresenter` or its own loader |
+| `activity` | 48 | 20 | Acceptable — thin wrapper |
+| `facts_seen_in_recent_recalls` | 418 | 20 | Move into `Dashboard::FactsQuery` |
+| `collect_configured_hook_types` | 739 | 19 | Move into `Dashboard::Health` |
+| `serialize_recall_fact` | 545 | 19 | Move into `Dashboard::RecallQuery` |
+| `health` | 14 | 18 | Becomes 3-liner after `Dashboard::Health` extraction |
+| `reject_fact` | 294 | 16 | Acceptable — public surface |
+
+### `sweep/maintenance.rb` (456 LOC)
+
+| Method | Line | Size | Action |
+|---|---|---|---|
+| `dedupe_open_conflicts` | 273 | 58 | Extract per-group `resolve_duplicate_group` helper |
+| `restore_multi_value_supersessions` | 185 | 57 | Already documented; could extract `compute_restore_decisions` |
+| `dedupe_multi_value_facts` | 58 | 34 | Acceptable — well-bounded transactional op |
+| `reclassify_references` | 340 | 26 | Acceptable |
+| `prune_old_content` | 130 | 16 | Acceptable |
+
+### `store/sqlite_store.rb` (584 LOC)
+
+| Method | Line | Size | Notes |
+|---|---|---|---|
+| `upsert_content_item` | 193 | 27 | 11 kwargs (carried #8) |
+| `reject_fact` | 410 | 25 | Conflict resolution in transaction |
+| `insert_fact` | 332 | 22 | Many optional fields |
+| `upsert_moment_feedback` | 123 | 21 | New — transaction with retry |
+| `update_fact` | 373 | 19 | Generic update via allowed-keys |
+
+---
+
+## Historical Reviews
+
+Earlier reviews (Jan 29, Feb 4, Mar 9, Mar 19) tracked the codebase from ~8,000 → 12,239 LOC. Their highlights, preserved here:
+
+- **Jan 29 (initial)** — Identified Tools and Recall god-object risks; introduced first metrics baseline.
+- **Feb 4** — Carried-forward items #17–#25 (DateTime migration, command manager helper, release_connections polymorphism, provenance batch insert, result objects). All still low-priority and open.
+- **Mar 9** — Three files >500 LOC; bare rescue counted; vector index work landed.
+- **Mar 19** — Successful refactor wave: `RetryHandler` + `SchemaManager` extracted from `SQLiteStore` (547 → 386); `Tools` reduced to 104-line dispatcher with 6 handler modules; `Recall` to 94-line facade. **Established the module-inclusion pattern** that has been reused successfully for LLMCache, MetricsAggregator, and the dashboard subsystems.
+
+The 2026-04-22 review absorbed the 39% codebase growth (+4,775 LOC) without correctness regressions and resolved its top two watch-items (`SQLiteStore` regrowth, dashboard test coverage). It left `Dashboard::API` extraction as a medium-priority watch item — which the present review (2026-04-28) escalates to high-priority based on the 180-LOC regression in 6 days.
 
 ---
 
-**Next review:** After Recall strategy pattern refactoring or SQLiteStore extraction
+**Next review:** After #31 (Dashboard::API extraction) lands, or pre-0.10.0 release tag.