pi-mega-compact 0.21.1 → 0.21.2

package/dist/config.js

@@ -116,8 +116,19 @@ export const NEW_UI = () => ragEnabled("MEGACOMPACT_NEW_UI");
 // default (0.12) keeps recall permissive within a repo while still rejecting
 // effectively-unrelated hits. Call-time read so tests can set the env per-test.
 // ---------------------------------------------------------------------------
-/** Same-repo recall cosine floor: top winner must be >= this to be injected. */
-export const RECALL_MIN_COSINE = () => Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+/**
+ * Same-repo recall cosine floor: top winner must be >= this to be injected.
+ *
+ * E1 follow-up (PR #18 review): NaN-safe + clamped to [0,1]. A typo'd env var
+ * yielded NaN before; `cosine < NaN` is false, which disabled gate 1 entirely
+ * (every candidate passed). Non-finite falls back to 0.12; out-of-range clamps.
+ */
+export const RECALL_MIN_COSINE = () => {
+    const n = Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+    if (!Number.isFinite(n))
+        return 0.12;
+    return Math.min(1, Math.max(0, n));
+};
 // ---------------------------------------------------------------------------
 // Vector-cortex flags + breaker constants (VC0A+). Positive sprint flags,
 // default ON, `=0`/`_DISABLED` off. Re-exported from src/config/vector-cortex.ts

package/dist/dedup/l1-minhash.js

@@ -19,6 +19,7 @@ export const SHINGLE_SIZE = 5; // char 5-grams
 const MAX_SHINGLES = 50_000; // QA #7/#15 complexity cap
 const SEED = 0xdeadbeef;
 const P = 2147483647; // 2^31 - 1, Mersenne prime
+const PBigInt = 2147483647n; // BigInt twin for overflow-safe modular reduction
 /** Per-index universal-hashing coefficients, derived deterministically from SEED. */
 function coeffA(i) {
     return (SEED + i * 2 + 1) % P;
@@ -67,10 +68,10 @@ export function minhashSignature(text) {
         const b = coeffB(i);
         let min = P;
         for (const x of grams) {
-            // (a*x + b) mod p — use Number math; a,x < 2^31 so a*x < 2^62, within
-            // double-precision integer range (2^53) only if reduced; reduce a*x first.
-            const ax = (a * (x % P)) % P;
-            const h = (ax + b) % P;
+            // (a*x + b) mod p. a, x < 2^31 so a*x < 2^62 — EXCEEDS 2^53. The naive
+            // (a*(x%P))%P loses precision (verified: a=x=p-1 → lossy 2147483644 vs
+            // exact 1). BigInt is correct + cheap (~5ms per signature).
+            const h = Number((BigInt(a) * BigInt(x % P) + BigInt(b)) % PBigInt);
             if (h < min)
                 min = h;
         }

package/dist/extensions/mega-compact.js

@@ -83,6 +83,16 @@ export default function (pi) {
         console.warn("[mega-compact] MEGACOMPACT_POISONED_REPEAT_THRESHOLD must be >= 1; using default 3");
         config.poisonedContextRepeatThreshold = 3;
     }
+    // E1: validate similarity thresholds — NaN or out of range silently disables
+    // recall dedup (anything >= NaN is false). envFlag guards NaN; clamp the rest.
+    if (!(config.dedupSim > 0 && config.dedupSim <= 1)) {
+        console.warn("[mega-compact] MEGACOMPACT_DEDUP_SIM must be in (0,1]; using default 0.9");
+        config.dedupSim = 0.9;
+    }
+    if (!(config.crossRepoCosine >= 0 && config.crossRepoCosine <= 1)) {
+        console.warn("[mega-compact] MEGACOMPACT_CROSSREPO_COSINE must be in [0,1]; using default 0.9");
+        config.crossRepoCosine = 0.9;
+    }
     const runtime = new MegaRuntime(config);
     registerEventHandlers(pi, runtime, config);
     registerCommands(pi, runtime, config);

package/dist/extensions/mega-config.js

@@ -171,7 +171,7 @@ export function loadConfig() {
         advisoryChannel: envBool("MEGACOMPACT_ADVISORY_CHANNEL", true),
         autoPctTrigger,
         autoInlineK: envFlag("MEGACOMPACT_AUTO_INLINE_K", 3),
-        dedupSim: Number(process.env.MEGACOMPACT_DEDUP_SIM ?? "0.9"),
+        dedupSim: envFlag("MEGACOMPACT_DEDUP_SIM", 0.9),
         raptorEnabled: envBool("MEGACOMPACT_RAPTOR_ENABLED", true),
         legacyDurableTrim: envBool("MEGACOMPACT_LEGACY_DURABLE_TRIM", false),
         dbMirror: envBool("MEGACOMPACT_DB_MIRROR", false),
@@ -180,13 +180,16 @@ export function loadConfig() {
         turnsDbEnabled: envBool("MEGACOMPACT_TURNS_DB", true),
         autoWikiEnabled: envBool("MEGACOMPACT_AUTO_WIKI", true),
         crossRepoEnabled: envBool("MEGACOMPACT_CROSSREPO_ENABLED", true),
-        crossRepoCosine: Number(process.env.MEGACOMPACT_CROSSREPO_COSINE ?? "0.90"),
+        crossRepoCosine: envFlag("MEGACOMPACT_CROSSREPO_COSINE", 0.9),
         // 3WF-3: SAME-repo recall cosine floor applied by the 3-source validator to
         // the top winner. SEPARATE from crossRepoCosine (S17, default 0.90, stricter
         // and cross-repo only). This same-repo floor is permissive by default (0.12)
         // so recall still surfaces loosely-relevant within-repo context while
         // rejecting effectively-unrelated hits. Mirrors src/config.ts RECALL_MIN_COSINE.
-        recallMinCosine: Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12"),
+        // E1 follow-up (PR #18 review): envFlag (Number.isFinite-guarded) like the
+        // dedupSim/crossRepoCosine fix in PR #18 — a typo'd env var must fall back
+        // to 0.12, not yield NaN.
+        recallMinCosine: envFlag("MEGACOMPACT_RECALL_MIN_COSINE", 0.12),
         memoryAutoReview: envBool("MEGACOMPACT_MEMORY_AUTO_REVIEW", true),
         memoryReviewInterval: envFlag("MEGACOMPACT_MEMORY_REVIEW_INTERVAL", 10),
         recallMaxTokens: envFlag("MEGACOMPACT_RECALL_MAX_TOKENS", 1500),

package/dist/src/config.js

@@ -116,8 +116,19 @@ export const NEW_UI = () => ragEnabled("MEGACOMPACT_NEW_UI");
 // default (0.12) keeps recall permissive within a repo while still rejecting
 // effectively-unrelated hits. Call-time read so tests can set the env per-test.
 // ---------------------------------------------------------------------------
-/** Same-repo recall cosine floor: top winner must be >= this to be injected. */
-export const RECALL_MIN_COSINE = () => Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+/**
+ * Same-repo recall cosine floor: top winner must be >= this to be injected.
+ *
+ * E1 follow-up (PR #18 review): NaN-safe + clamped to [0,1]. A typo'd env var
+ * yielded NaN before; `cosine < NaN` is false, which disabled gate 1 entirely
+ * (every candidate passed). Non-finite falls back to 0.12; out-of-range clamps.
+ */
+export const RECALL_MIN_COSINE = () => {
+    const n = Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+    if (!Number.isFinite(n))
+        return 0.12;
+    return Math.min(1, Math.max(0, n));
+};
 // ---------------------------------------------------------------------------
 // Vector-cortex flags + breaker constants (VC0A+). Positive sprint flags,
 // default ON, `=0`/`_DISABLED` off. Re-exported from src/config/vector-cortex.ts

package/dist/src/dedup/l1-minhash.js

@@ -19,6 +19,7 @@ export const SHINGLE_SIZE = 5; // char 5-grams
 const MAX_SHINGLES = 50_000; // QA #7/#15 complexity cap
 const SEED = 0xdeadbeef;
 const P = 2147483647; // 2^31 - 1, Mersenne prime
+const PBigInt = 2147483647n; // BigInt twin for overflow-safe modular reduction
 /** Per-index universal-hashing coefficients, derived deterministically from SEED. */
 function coeffA(i) {
     return (SEED + i * 2 + 1) % P;
@@ -67,10 +68,10 @@ export function minhashSignature(text) {
         const b = coeffB(i);
         let min = P;
         for (const x of grams) {
-            // (a*x + b) mod p — use Number math; a,x < 2^31 so a*x < 2^62, within
-            // double-precision integer range (2^53) only if reduced; reduce a*x first.
-            const ax = (a * (x % P)) % P;
-            const h = (ax + b) % P;
+            // (a*x + b) mod p. a, x < 2^31 so a*x < 2^62 — EXCEEDS 2^53. The naive
+            // (a*(x%P))%P loses precision (verified: a=x=p-1 → lossy 2147483644 vs
+            // exact 1). BigInt is correct + cheap (~5ms per signature).
+            const h = Number((BigInt(a) * BigInt(x % P) + BigInt(b)) % PBigInt);
             if (h < min)
                 min = h;
         }

package/dist/src/recall/validator.js

@@ -76,7 +76,14 @@ export function validateRecall(winners, opts, store) {
                 // No comparable cosine available => cannot clear a cosine gate.
                 continue;
             }
-            if (cosine < floor)
+            // E1 follow-up (PR #18 review): NaN/Infinity must NEVER clear the floor.
+            // `NaN < floor` is false, so an unguarded comparison lets a NaN cosine
+            // PASS gate 1 and inject — one NaN source poisons the whole 3WF-3
+            // quorum. Reject non-finite scores explicitly; the candidate is skipped
+            // and, if all fail, the provenance floor ("no recall") is returned —
+            // never a zero-score injection. The default TrigramEmbedder cannot
+            // produce NaN (zero-norm guard), but a BYO localhost embedder can.
+            if (!Number.isFinite(cosine) || cosine < floor)
                 continue;
             // Gate 2: not already resident in the live window.
             if (liveVecs.length > 0) {

package/dist/src/recall/vote.js

@@ -31,17 +31,36 @@ import { listCheckpoints } from "../store/sqlite.js";
 import { computeContentDigest } from "../dedup/digest.js";
 import { recallRawHits } from "./readonly.js";
 import { Logger } from "../log.js";
-/** Per-source normalization: map raw scores to 0..1 via min-max within source. */
-function normalizeScores(scores) {
+/**
+ * Per-source normalization: map raw scores to 0..1 via min-max within source.
+ *
+ * E1 follow-up (PR #18 review): non-finite scores (NaN/±Infinity) are DROPPED
+ * before the min/max fold — a single NaN silently propagates through Math.min/
+ * max and turns EVERY normalized score of that source into NaN (verified),
+ * poisoning the whole 3-source quorum. Dropping the bad entry degrades that
+ * source gracefully instead. Exported so the guard is unit-testable directly.
+ */
+export function normalizeScores(scores) {
     const map = new Map();
-    if (scores.length === 0)
-        return map;
-    const min = Math.min(...scores);
-    const max = Math.max(...scores);
-    const span = max - min;
+    const finite = [];
     scores.forEach((s, i) => {
-        map.set(i, span === 0 ? 1 : (s - min) / span);
+        if (Number.isFinite(s))
+            finite.push({ idx: i, s });
     });
+    if (finite.length === 0)
+        return map;
+    let min = Infinity;
+    let max = -Infinity;
+    for (const { s } of finite) {
+        if (s < min)
+            min = s;
+        if (s > max)
+            max = s;
+    }
+    const span = max - min;
+    for (const { idx, s } of finite) {
+        map.set(idx, span === 0 ? 1 : (s - min) / span);
+    }
     return map;
 }
 /**
@@ -130,7 +149,13 @@ export function voteRecall(opts, store) {
         const norm = perSource.find((p) => p.name === src.name).norm;
         const bestByCp = new Map();
         src.cands.forEach((c, i) => {
-            const n = norm.get(i) ?? 0;
+            // E1 follow-up: normalizeScores dropped non-finite scores; such a hit is
+            // NOT a valid nomination — skip it entirely instead of defaulting it to
+            // 0 (which would still name the checkpoint and let it rank last into the
+            // fallback ranking).
+            const n = norm.get(i);
+            if (n === undefined)
+                return;
             const prev = bestByCp.get(c.checkpointId);
             if (prev === undefined || n > prev)
                 bestByCp.set(c.checkpointId, n);

package/dist/src/store/sqlite/fts5-search.js

@@ -87,7 +87,12 @@ export function hydrateFts5Hits(hits, sessionId, stateDir) {
     const out = [];
     for (const h of hits) {
         const cp = cpMap.get(h.id);
-        if (cp)
+        // H1 follow-up (PR #18 review): exclude SemDeDup-'removed' rows. The FTS5
+        // index is NOT pruned by vectorSemDedup, so a removed row can still MATCH;
+        // hydrating it would let the fts5 voter NAME a dead checkpoint (a removed
+        // row plus a recency vote is a 2/3 agreement that passes the validator).
+        // Mirrors vectorSearch's read-time filter.
+        if (cp && cp.dedupStatus !== "removed")
             out.push({ checkpointId: h.id, score: h.score, summary: cp.summary });
     }
     return out;

package/dist/src/vector-read.js

@@ -72,7 +72,10 @@ export function vectorDedupe(store, sessionId, regionHashOrText, isText = false)
     const state = loadSessionState(sid, stateDir);
     if (state.storedRegionHashes.includes(hash))
         return true;
-    return listCheckpoints(sid, stateDir).some((c) => c.regionHash === hash);
+    // H1 follow-up (PR #18 review): a SemDeDup-'removed' row's regionHash must not
+    // report "already represented" — its content is excluded from recall, so the
+    // incoming region is NOT deduplicated in any retrievable sense.
+    return listCheckpoints(sid, stateDir).some((c) => c.dedupStatus !== "removed" && c.regionHash === hash);
 }
 // ---------------------------------------------------------------------------
 // Injection tracking
@@ -115,7 +118,8 @@ export function vectorTopSimilar(store, sessionId, n) {
     const ordered = [...checkpoints].sort((a, b) => a.checkpointId.localeCompare(b.checkpointId));
     const current = ordered[ordered.length - 1];
     const scored = ordered
-        .filter((cp) => cp.checkpointId !== current.checkpointId)
+        .filter((cp) => cp.checkpointId !== current.checkpointId &&
+        cp.dedupStatus !== "removed")
         .map((cp) => ({
         checkpoint: cp,
         score: cosineSimilarity(current.embedding, cp.embedding),

package/dist/src/vectorStore/add.js

@@ -40,7 +40,11 @@ export function addCheckpoint(store, input) {
     const t0 = Date.now();
     const sessionId = normalizeSessionId(input.sessionId);
     const regionHash = computeRegionHash(input.regionText);
-    const all = listCheckpoints(sessionId, store.stateDir);
+    // H1: exclude SemDeDup-'removed' rows from dedup matching. search() already
+    // filters these, but add() did NOT — so an L0/L1/L2 match against a previously
+    // removed duplicate would upsertCheckpoint it back to active, resurrecting it
+    // into recall and defeating SemDeDup.
+    const all = listCheckpoints(sessionId, store.stateDir).filter((cp) => cp.dedupStatus !== "removed");
     // Honest "tokens saved" base for this region. For a deduped add the whole
     // original region is discarded (nothing new stored); for a new checkpoint
     // we persist (orig − stored). Falls back to stored when orig is unknown.

package/extensions/mega-compact.ts

@@ -98,6 +98,16 @@ export default function (pi: ExtensionAPI) {
 		);
 		config.poisonedContextRepeatThreshold = 3;
 	}
+	// E1: validate similarity thresholds — NaN or out of range silently disables
+	// recall dedup (anything >= NaN is false). envFlag guards NaN; clamp the rest.
+	if (!(config.dedupSim > 0 && config.dedupSim <= 1)) {
+		console.warn("[mega-compact] MEGACOMPACT_DEDUP_SIM must be in (0,1]; using default 0.9");
+		config.dedupSim = 0.9;
+	}
+	if (!(config.crossRepoCosine >= 0 && config.crossRepoCosine <= 1)) {
+		console.warn("[mega-compact] MEGACOMPACT_CROSSREPO_COSINE must be in [0,1]; using default 0.9");
+		config.crossRepoCosine = 0.9;
+	}
 	const runtime = new MegaRuntime(config);
 	registerEventHandlers(pi, runtime, config);
 	registerCommands(pi, runtime, config);

package/extensions/mega-config.ts

@@ -207,7 +207,7 @@ export function loadConfig(): MegaConfig {
 		advisoryChannel: envBool("MEGACOMPACT_ADVISORY_CHANNEL", true),
 		autoPctTrigger,
 		autoInlineK: envFlag("MEGACOMPACT_AUTO_INLINE_K", 3),
-		dedupSim: Number(process.env.MEGACOMPACT_DEDUP_SIM ?? "0.9"),
+		dedupSim: envFlag("MEGACOMPACT_DEDUP_SIM", 0.9),
 		raptorEnabled: envBool("MEGACOMPACT_RAPTOR_ENABLED", true),
 		legacyDurableTrim: envBool("MEGACOMPACT_LEGACY_DURABLE_TRIM", false),
 		dbMirror: envBool("MEGACOMPACT_DB_MIRROR", false),
@@ -216,13 +216,16 @@ export function loadConfig(): MegaConfig {
 		turnsDbEnabled: envBool("MEGACOMPACT_TURNS_DB", true),
 		autoWikiEnabled: envBool("MEGACOMPACT_AUTO_WIKI", true),
 		crossRepoEnabled: envBool("MEGACOMPACT_CROSSREPO_ENABLED", true),
-		crossRepoCosine: Number(process.env.MEGACOMPACT_CROSSREPO_COSINE ?? "0.90"),
+		crossRepoCosine: envFlag("MEGACOMPACT_CROSSREPO_COSINE", 0.9),
 		// 3WF-3: SAME-repo recall cosine floor applied by the 3-source validator to
 		// the top winner. SEPARATE from crossRepoCosine (S17, default 0.90, stricter
 		// and cross-repo only). This same-repo floor is permissive by default (0.12)
 		// so recall still surfaces loosely-relevant within-repo context while
 		// rejecting effectively-unrelated hits. Mirrors src/config.ts RECALL_MIN_COSINE.
-		recallMinCosine: Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12"),
+		// E1 follow-up (PR #18 review): envFlag (Number.isFinite-guarded) like the
+		// dedupSim/crossRepoCosine fix in PR #18 — a typo'd env var must fall back
+		// to 0.12, not yield NaN.
+		recallMinCosine: envFlag("MEGACOMPACT_RECALL_MIN_COSINE", 0.12),
 		memoryAutoReview: envBool("MEGACOMPACT_MEMORY_AUTO_REVIEW", true),
 		memoryReviewInterval: envFlag("MEGACOMPACT_MEMORY_REVIEW_INTERVAL", 10),
 		recallMaxTokens: envFlag("MEGACOMPACT_RECALL_MAX_TOKENS", 1500),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.21.1",
+  "version": "0.21.2",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",

package/src/config.ts

@@ -152,9 +152,18 @@ export const NEW_UI = (): boolean => ragEnabled("MEGACOMPACT_NEW_UI");
 // effectively-unrelated hits. Call-time read so tests can set the env per-test.
 // ---------------------------------------------------------------------------
 
-/** Same-repo recall cosine floor: top winner must be >= this to be injected. */
-export const RECALL_MIN_COSINE = (): number =>
-	Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+/**
+ * Same-repo recall cosine floor: top winner must be >= this to be injected.
+ *
+ * E1 follow-up (PR #18 review): NaN-safe + clamped to [0,1]. A typo'd env var
+ * yielded NaN before; `cosine < NaN` is false, which disabled gate 1 entirely
+ * (every candidate passed). Non-finite falls back to 0.12; out-of-range clamps.
+ */
+export const RECALL_MIN_COSINE = (): number => {
+	const n = Number(process.env.MEGACOMPACT_RECALL_MIN_COSINE ?? "0.12");
+	if (!Number.isFinite(n)) return 0.12;
+	return Math.min(1, Math.max(0, n));
+};
 
 // ---------------------------------------------------------------------------
 // Vector-cortex flags + breaker constants (VC0A+). Positive sprint flags,

package/src/dedup/l1-minhash.ts

@@ -21,6 +21,7 @@ export const SHINGLE_SIZE = 5; // char 5-grams
 const MAX_SHINGLES = 50_000; // QA #7/#15 complexity cap
 const SEED = 0xdeadbeef;
 const P = 2147483647; // 2^31 - 1, Mersenne prime
+const PBigInt = 2147483647n; // BigInt twin for overflow-safe modular reduction
 
 /** Per-index universal-hashing coefficients, derived deterministically from SEED. */
 function coeffA(i: number): number {
@@ -69,10 +70,10 @@ export function minhashSignature(text: string): number[] {
     const b = coeffB(i);
     let min = P;
     for (const x of grams) {
-      // (a*x + b) mod p — use Number math; a,x < 2^31 so a*x < 2^62, within
-      // double-precision integer range (2^53) only if reduced; reduce a*x first.
-      const ax = (a * (x % P)) % P;
-      const h = (ax + b) % P;
+      // (a*x + b) mod p. a, x < 2^31 so a*x < 2^62 — EXCEEDS 2^53. The naive
+      // (a*(x%P))%P loses precision (verified: a=x=p-1 → lossy 2147483644 vs
+      // exact 1). BigInt is correct + cheap (~5ms per signature).
+      const h = Number((BigInt(a) * BigInt(x % P) + BigInt(b)) % PBigInt);
       if (h < min) min = h;
     }
     sig[i] = min;

package/src/recall/validator.ts

@@ -113,7 +113,14 @@ export function validateRecall(
 				// No comparable cosine available => cannot clear a cosine gate.
 				continue;
 			}
-			if (cosine < floor) continue;
+			// E1 follow-up (PR #18 review): NaN/Infinity must NEVER clear the floor.
+			// `NaN < floor` is false, so an unguarded comparison lets a NaN cosine
+			// PASS gate 1 and inject — one NaN source poisons the whole 3WF-3
+			// quorum. Reject non-finite scores explicitly; the candidate is skipped
+			// and, if all fail, the provenance floor ("no recall") is returned —
+			// never a zero-score injection. The default TrigramEmbedder cannot
+			// produce NaN (zero-norm guard), but a BYO localhost embedder can.
+			if (!Number.isFinite(cosine) || cosine < floor) continue;
 
 			// Gate 2: not already resident in the live window.
 			if (liveVecs.length > 0) {

package/src/recall/vote.ts

@@ -46,16 +46,32 @@ export interface VoteOptions {
 	recencyCount?: number;
 }
 
-/** Per-source normalization: map raw scores to 0..1 via min-max within source. */
-function normalizeScores(scores: number[]): Map<number, number> {
+/**
+ * Per-source normalization: map raw scores to 0..1 via min-max within source.
+ *
+ * E1 follow-up (PR #18 review): non-finite scores (NaN/±Infinity) are DROPPED
+ * before the min/max fold — a single NaN silently propagates through Math.min/
+ * max and turns EVERY normalized score of that source into NaN (verified),
+ * poisoning the whole 3-source quorum. Dropping the bad entry degrades that
+ * source gracefully instead. Exported so the guard is unit-testable directly.
+ */
+export function normalizeScores(scores: number[]): Map<number, number> {
 	const map = new Map<number, number>();
-	if (scores.length === 0) return map;
-	const min = Math.min(...scores);
-	const max = Math.max(...scores);
-	const span = max - min;
+	const finite: { idx: number; s: number }[] = [];
 	scores.forEach((s, i) => {
-		map.set(i, span === 0 ? 1 : (s - min) / span);
+		if (Number.isFinite(s)) finite.push({ idx: i, s });
 	});
+	if (finite.length === 0) return map;
+	let min = Infinity;
+	let max = -Infinity;
+	for (const { s } of finite) {
+		if (s < min) min = s;
+		if (s > max) max = s;
+	}
+	const span = max - min;
+	for (const { idx, s } of finite) {
+		map.set(idx, span === 0 ? 1 : (s - min) / span);
+	}
 	return map;
 }
 
@@ -150,7 +166,12 @@ export function voteRecall(opts: VoteOptions, store: VectorStore): VoteResult {
 		const norm = perSource.find((p) => p.name === src.name)!.norm;
 		const bestByCp = new Map<string, number>();
 		src.cands.forEach((c, i) => {
-			const n = norm.get(i) ?? 0;
+			// E1 follow-up: normalizeScores dropped non-finite scores; such a hit is
+			// NOT a valid nomination — skip it entirely instead of defaulting it to
+			// 0 (which would still name the checkpoint and let it rank last into the
+			// fallback ranking).
+			const n = norm.get(i);
+			if (n === undefined) return;
 			const prev = bestByCp.get(c.checkpointId);
 			if (prev === undefined || n > prev) bestByCp.set(c.checkpointId, n);
 			seenIds.add(c.checkpointId);

package/src/store/sqlite/fts5-search.ts

@@ -124,7 +124,13 @@ export function hydrateFts5Hits(
 	const out: HydratedFts5Hit[] = [];
 	for (const h of hits) {
 		const cp = cpMap.get(h.id);
-		if (cp) out.push({ checkpointId: h.id, score: h.score, summary: cp.summary });
+		// H1 follow-up (PR #18 review): exclude SemDeDup-'removed' rows. The FTS5
+		// index is NOT pruned by vectorSemDedup, so a removed row can still MATCH;
+		// hydrating it would let the fts5 voter NAME a dead checkpoint (a removed
+		// row plus a recency vote is a 2/3 agreement that passes the validator).
+		// Mirrors vectorSearch's read-time filter.
+		if (cp && cp.dedupStatus !== "removed")
+			out.push({ checkpointId: h.id, score: h.score, summary: cp.summary });
 	}
 	return out;
 }

package/src/vector-read.ts

@@ -100,8 +100,11 @@ export function vectorDedupe(
     : regionHashOrText;
   const state = loadSessionState(sid, stateDir);
   if (state.storedRegionHashes.includes(hash)) return true;
+  // H1 follow-up (PR #18 review): a SemDeDup-'removed' row's regionHash must not
+  // report "already represented" — its content is excluded from recall, so the
+  // incoming region is NOT deduplicated in any retrievable sense.
   return listCheckpoints(sid, stateDir).some(
-    (c) => c.regionHash === hash,
+    (c) => c.dedupStatus !== "removed" && c.regionHash === hash,
   );
 }
 
@@ -157,7 +160,11 @@ export function vectorTopSimilar(store: VectorStore, sessionId: string, n: numbe
   const current = ordered[ordered.length - 1];
 
   const scored: SearchHit[] = ordered
-    .filter((cp) => cp.checkpointId !== current.checkpointId)
+    .filter(
+      (cp) =>
+        cp.checkpointId !== current.checkpointId &&
+        cp.dedupStatus !== "removed", // H1 follow-up: mirror vectorSearch's filter
+    )
     .map((cp) => ({
       checkpoint: cp,
       score: cosineSimilarity(current.embedding, cp.embedding),

package/src/vectorStore/add.ts

@@ -54,7 +54,13 @@ export function addCheckpoint(store: VectorStore, input: AddInput): AddResult {
 	const t0 = Date.now();
 	const sessionId = normalizeSessionId(input.sessionId);
 	const regionHash = computeRegionHash(input.regionText);
-	const all = listCheckpoints(sessionId, store.stateDir);
+	// H1: exclude SemDeDup-'removed' rows from dedup matching. search() already
+	// filters these, but add() did NOT — so an L0/L1/L2 match against a previously
+	// removed duplicate would upsertCheckpoint it back to active, resurrecting it
+	// into recall and defeating SemDeDup.
+	const all = listCheckpoints(sessionId, store.stateDir).filter(
+		(cp) => cp.dedupStatus !== "removed",
+	);
 	// Honest "tokens saved" base for this region. For a deduped add the whole
 	// original region is discarded (nothing new stored); for a new checkpoint
 	// we persist (orig − stored). Falls back to stored when orig is unknown.