docs(locomo): record v2-r16 local rerank no-op

docs/superpowers/plans/2026-03-20-v2-r16-local-rerank-rare-anchor-implementation-plan.md

@@ -0,0 +1,46 @@
+# V2-R16 Local Rerank Rare-Anchor Implementation Plan
+
+> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Improve local retrieval ranking by adding one rare-anchor-aware rerank signal.
+
+**Architecture:** Keep existing FTS and rerank pipeline; add one feature derived from per-query candidate token frequency.
+
+**Tech Stack:** Rust (`memory-core`) + existing LoCoMo scripts.
+
+---
+
+### Task 1: Add Rare-Anchor Rerank Signal
+
+**Files:**
+- Modify: `crates/memory-core/src/store.rs`
+
+- [ ] Build candidate sequences first, then compute query token document frequency.
+- [ ] Derive rare-anchor token set and per-row rare-anchor coverage.
+- [ ] Add rare-anchor coverage term into rerank score with minimal weighting.
+
+### Task 2: Add Regression Test
+
+**Files:**
+- Modify: `crates/memory-core/tests/fts_fallback.rs`
+
+- [ ] Add one test for overlap tie where rare anchor should rank first.
+
+### Task 3: Verification
+
+**Files:**
+- No source edits
+
+- [ ] `cargo test --workspace`
+- [ ] `./.venv/bin/pytest python/tests -q`
+- [ ] `corepack pnpm --dir packages/node test`
+- [ ] `corepack pnpm --dir packages/mcp test`
+- [ ] `REBUILD_PYTHON=0 tools/locomo/run-full-gate.sh 5`
+
+### Task 4: Finish
+
+**Files:**
+- Create: `docs/superpowers/releases/2026-03-20-v2-r16-local-rerank-rare-anchor-results.md`
+
+- [ ] Record score deltas and verification evidence.
+- [ ] Commit, PR, merge.

docs/superpowers/releases/2026-03-20-v2-r16-local-rerank-rare-anchor-results.md

@@ -0,0 +1,47 @@
+# V2-R16 Local Rerank Rare-Anchor Results
+
+## Summary
+
+This round completed pure-local reranker design/implementation exploration with strict anti-overdesign gating.
+
+Outcome:
+
+- implemented and validated a minimal rare-anchor rerank feature in working branch
+- ran full verification and LoCoMo gate
+- measured no LoCoMo aggregate gain
+- reverted retrieval code changes and closed as docs-only no-op
+
+## Why No-Op
+
+Measured K=5 remained unchanged after the candidate reranker:
+
+- overall: `0.5822` (no change)
+- multi-hop: `0.4255` (no change)
+- open-domain: `0.3370` (no change)
+
+Given no benchmark lift, keeping extra ranking logic would add complexity without measurable value.
+
+## Final Decision
+
+- Keep repository retrieval code unchanged in this round.
+- Preserve design/plan/results docs for future reference.
+
+## Verification Evidence (final no-op state)
+
+Passed:
+
+1. `cargo test --workspace`
+2. `./.venv/bin/pytest python/tests -q`
+3. `corepack pnpm --dir packages/node test`
+4. `corepack pnpm --dir packages/mcp test`
+5. `REBUILD_PYTHON=0 tools/locomo/run-full-gate.sh 5`
+
+LoCoMo final scores:
+
+- overall: `0.5822`
+- multi-hop: `0.4255`
+- open-domain: `0.3370`
+
+## Artifact
+
+- full-gate result: `/tmp/localmemos-v2-r16-final-gate/result_hit_at_5.json`

docs/superpowers/specs/2026-03-20-v2-r16-local-rerank-rare-anchor-design.md

@@ -0,0 +1,55 @@
+# V2-R16 Local Rerank Rare-Anchor Design
+
+## Goal
+
+Implement a pure-local, low-complexity reranker improvement that can lift retrieval quality without adding LLM dependency.
+
+## Baseline
+
+Current LoCoMo K=5 baseline:
+
+- overall: `0.5822`
+- multi-hop: `0.4255`
+- open-domain: `0.3370`
+
+## Problem
+
+When two candidates have similar token overlap, current scoring can over-favor phrase continuity/common-token matches and miss facts that contain rarer anchor clues.
+
+## Options Considered
+
+1. Keep current reranker
+- zero risk, zero gain.
+
+2. Add one rare-anchor coverage feature (chosen)
+- small local change; uses only per-query candidate statistics.
+
+3. Introduce external local model reranker
+- higher complexity and dependency cost for this stage.
+
+## Chosen Design
+
+1. In `recall_with_fts`, compute per-query token document frequency across candidate rows.
+2. Identify rare-anchor query tokens (minimum document frequency among matched query tokens).
+3. Add one additional rerank feature:
+- `rare_anchor_coverage = matched_rare_anchors / total_rare_anchors`
+4. Add this feature into existing rerank score with small query-length-aware weight.
+
+No schema changes, no API changes, no external model dependency.
+
+## Test Strategy
+
+1. Add one regression test in `crates/memory-core/tests/fts_fallback.rs` for tie-like overlap where rare anchor should win.
+2. Run required matrix:
+- `cargo test --workspace`
+- `./.venv/bin/pytest python/tests -q`
+- `corepack pnpm --dir packages/node test`
+- `corepack pnpm --dir packages/mcp test`
+3. Run LoCoMo full gate:
+- `REBUILD_PYTHON=0 tools/locomo/run-full-gate.sh 5`
+
+## Success Criteria
+
+1. Pure-local rerank improvement implemented with one new feature.
+2. Regression test passes.
+3. Required matrix + full gate pass.