pi-mega-compact 0.11.5 → 0.11.7

package/dist/src/recallMetrics.js

@@ -0,0 +1,258 @@
+/**
+ * recallMetrics.ts — S45 CRAG quality metrics for recall evaluation.
+ *
+ * Four metrics evaluate the quality of a set of recall hits:
+ *   - relevance:     average cosine similarity (SearchHit.score) across hits
+ *   - coverage:      fraction of query sub-terms found in at least one hit summary
+ *   - diversity:     average pairwise cosine distance among top-K embeddings (MMR-style)
+ *   - specificity:   inverse document-frequency proxy via average chunk length
+ *
+ * Pi-agnostic: no pi runtime types, no network calls (PREVENT-PI-004).
+ * All config is passed explicitly; defaults are documented but NOT hardcoded
+ * inside the metric functions (the four constants from S45A.5 are also received
+ * via config, not baked in).
+ *
+ * Best-effort: all functions handle degenerate inputs (empty hits, missing
+ * embeddings, zero query terms) and return sensible defaults instead of NaN.
+ */
+import { cosineSimilarity } from "./embedder.js";
+import { STOP_WORDS } from "./config/stopwords.js";
+/** Default config — all values marked "uncalibrated" unless CRAG_CALIBRATED=true. */
+export const DEFAULT_RECALL_QUALITY_CONFIG = {
+    minDiversity: 0.3,
+    minCoverage: 0.4,
+    minRelevance: 0.5,
+    minOverallScore: 0.4,
+    weights: { relevance: 0.35, coverage: 0.25, diversity: 0.25, specificity: 0.15 },
+    specificityDivisor: 300,
+    specificityOptimalMin: 100,
+    specificityOptimalMax: 500,
+    idfBroadenRatio: 0.5,
+    stopwords: STOP_WORDS,
+    calibrated: false,
+};
+// ---------------------------------------------------------------------------
+// Metric helpers (exported individually for unit testing)
+// ---------------------------------------------------------------------------
+/**
+ * S45A.2: Relevance — average of SearchHit.score across all hits.
+ * Scores are already cosine similarities in [0, 1].
+ * Empty hits → 0.
+ */
+export function computeRelevance(hits) {
+    if (hits.length === 0)
+        return 0;
+    let sum = 0;
+    for (const h of hits)
+        sum += h.score;
+    return sum / hits.length;
+}
+/**
+ * S45A.3: Coverage — fraction of query subterms present in at least one
+ * hit's `checkpoint.summary`. Tokenization removes stopwords and short terms.
+ * Zero query terms after filtering → coverage = 1.0 (trivially satisfied).
+ */
+export function computeCoverage(query, hits, stopwords = STOP_WORDS) {
+    const queryTerms = tokenizeTerms(query, stopwords);
+    if (queryTerms.length === 0)
+        return 1.0;
+    if (hits.length === 0)
+        return 0;
+    let foundCount = 0;
+    for (const qt of queryTerms) {
+        const found = hits.some((h) => {
+            const summary = h.checkpoint.summary ?? "";
+            return summary.toLowerCase().includes(qt);
+        });
+        if (found)
+            foundCount++;
+    }
+    return foundCount / queryTerms.length;
+}
+/**
+ * S45A.4: Diversity — average pairwise cosine distance (1 - similarity)
+ * among hit embeddings. Filtered to hits with non-empty embedding arrays.
+ *
+ * Edge cases:
+ *   - < 2 valid hits → 1.0 (nothing to compare)
+ *   - empty-embedding checkpoints are skipped (no NaN path)
+ */
+export function computeDiversity(hits) {
+    const valid = hits.filter((h) => Array.isArray(h.checkpoint.embedding) && h.checkpoint.embedding.length > 0);
+    if (valid.length < 2)
+        return 1.0;
+    let totalDist = 0;
+    let pairCount = 0;
+    for (let i = 0; i < valid.length; i++) {
+        for (let j = i + 1; j < valid.length; j++) {
+            const sim = cosineSimilarity(valid[i].checkpoint.embedding, valid[j].checkpoint.embedding);
+            // cosineSimilarity returns 0 for empty vectors, but we've filtered those out.
+            // NaN/Infinity could theoretically arise from degenerate floats; guard defensively.
+            const dist = Number.isFinite(sim) ? 1 - sim : 1.0;
+            totalDist += dist;
+            pairCount++;
+        }
+    }
+    return pairCount > 0 ? totalDist / pairCount : 1.0;
+}
+/**
+ * S45A.5: Specificity — average chunk length (tokenEstimate) as a proxy
+ * for information density.
+ *
+ * Formula (all constants from config, NOT hardcoded):
+ *   avg = mean(tokenEstimate across hits; hits missing tokenEstimate → 0)
+ *   If avg < optimalMin / 2 → 0 (too short to be specific)
+ *   base = avg / divisor
+ *   If avg > optimalMax:
+ *     ramp = 1 - (avg - optimalMax) / optimalMax  // linear to 0 at 2*optimalMax
+ *     capped = max(0, ramp)
+ *     specificity = min(base, capped)
+ *   else:
+ *     specificity = min(base, 1)
+ *
+ * Edge cases: empty hits → 1.0 (nothing to penalize).
+ */
+export function computeSpecificity(hits, config) {
+    if (hits.length === 0)
+        return 1.0;
+    const { specificityDivisor: divisor, specificityOptimalMin: optMin, specificityOptimalMax: optMax } = config;
+    let sum = 0;
+    for (const h of hits) {
+        sum += h.checkpoint.tokenEstimate ?? 0;
+    }
+    const avg = sum / hits.length;
+    // Below optMin/2 → too short to be specific
+    if (avg < optMin / 2)
+        return 0;
+    const base = divisor > 0 ? avg / divisor : 0;
+    const clamped = base > 1 ? 1 : base;
+    if (avg > optMax) {
+        // Linear ramp-down from optMax to 0 at 2*optMax
+        const ramp = optMax > 0 ? 1 - (avg - optMax) / optMax : 0;
+        const capped = ramp > 0 ? ramp : 0;
+        return clamped < capped ? clamped : capped;
+    }
+    return clamped;
+}
+// ---------------------------------------------------------------------------
+// Composite evaluation
+// ---------------------------------------------------------------------------
+/**
+ * Normalize a four-weight vector so its components sum to 1.0.
+ * If all weights are 0, return equal weights (0.25 each).
+ * Used defensively since users may set non-sum-1 weights via env.
+ */
+export function normalizeWeights(w) {
+    const sum = w.relevance + w.coverage + w.diversity + w.specificity;
+    if (sum === 0)
+        return { relevance: 0.25, coverage: 0.25, diversity: 0.25, specificity: 0.25 };
+    return {
+        relevance: w.relevance / sum,
+        coverage: w.coverage / sum,
+        diversity: w.diversity / sum,
+        specificity: w.specificity / sum,
+    };
+}
+/**
+ * S45A.6: Evaluate recall quality — compute all four metrics and return a
+ * composite result with pass/fail + recommendation.
+ */
+export function evaluateRecall(query, hits, config) {
+    const weights = normalizeWeights(config.weights);
+    const breakdown = {
+        relevance: computeRelevance(hits),
+        coverage: computeCoverage(query, hits, config.stopwords),
+        diversity: computeDiversity(hits),
+        specificity: computeSpecificity(hits, config),
+    };
+    const score = weights.relevance * breakdown.relevance +
+        weights.coverage * breakdown.coverage +
+        weights.diversity * breakdown.diversity +
+        weights.specificity * breakdown.specificity;
+    const pass = score >= config.minOverallScore &&
+        breakdown.relevance >= config.minRelevance &&
+        breakdown.coverage >= config.minCoverage &&
+        breakdown.diversity >= config.minDiversity;
+    const recommendation = buildRecommendation(breakdown, score, config);
+    return {
+        pass,
+        score,
+        breakdown,
+        recommendation,
+        uncalibrated: !config.calibrated,
+        weights,
+    };
+}
+// ---------------------------------------------------------------------------
+// Recommendation builder
+// ---------------------------------------------------------------------------
+/**
+ * Build a human-readable recommendation based on which metric is lowest.
+ * Uses env-overridable thresholds from config (not hardcoded).
+ */
+function buildRecommendation(breakdown, score, config) {
+    // Check each metric against its threshold (lowest first for specificity).
+    const issues = [];
+    if (breakdown.diversity < config.minDiversity) {
+        issues.push("chunks are too similar — consider expanding K or using MMR");
+    }
+    if (breakdown.coverage < config.minCoverage) {
+        issues.push("query terms not well covered — consider query expansion");
+    }
+    if (breakdown.relevance < config.minRelevance) {
+        issues.push("chunks may be irrelevant — consider stricter threshold");
+    }
+    if (score < config.minOverallScore) {
+        issues.push("quality is low — recommend re-retrieval with expanded query");
+    }
+    return issues.length > 0 ? issues.join("; ") : null;
+}
+/**
+ * Compute recall quality metrics for a set of search hits.
+ *
+ * Pi-agnostic: takes query text and SearchHit results directly, no pi types.
+ * The `config` parameter is optional — when omitted, DEFAULT_RECALL_QUALITY_CONFIG
+ * is used (labelled uncalibrated).
+ *
+ * This is the primary export for external callers (e.g., recall.ts integration).
+ */
+export function computeRecallMetrics(query, results, config) {
+    const effectiveConfig = {
+        ...DEFAULT_RECALL_QUALITY_CONFIG,
+        ...config,
+        // Merge weights shallowly if both sides exist
+        weights: {
+            ...DEFAULT_RECALL_QUALITY_CONFIG.weights,
+            ...(config?.weights ?? {}),
+        },
+    };
+    const result = evaluateRecall(query, results, effectiveConfig);
+    return {
+        breakdown: result.breakdown,
+        score: result.score,
+        pass: result.pass,
+        recommendation: result.recommendation,
+        uncalibrated: result.uncalibrated,
+    };
+}
+// ---------------------------------------------------------------------------
+// Internal helpers
+// ---------------------------------------------------------------------------
+/**
+ * Tokenize text into lowercased alphanumeric terms, filtering stop words
+ * and terms shorter than 3 characters.
+ */
+function tokenizeTerms(text, stopwords) {
+    return text
+        .toLowerCase()
+        .split(/[^a-z0-9]+/)
+        .filter((t) => t.length >= 3 && !stopwords.has(t));
+}
+/**
+ * Estimate block tokens for a piece of text (rough approximation).
+ * Used by the specificity metric fallback when tokenEstimate is missing.
+ */
+export function estimateBlockTokens(text) {
+    // ~4 chars per token for English text
+    return Math.ceil(text.length / 4);
+}

package/dist/src/store/sqlite/cache-stripes.js

@@ -0,0 +1,133 @@
+/**
+ * src/store/sqlite/cache-stripes.ts — Cache stripe distribution reader (A3, PLAN_V2 Phase 4).
+ *
+ * Reads from the cache_stripes table (schema.ts) to produce stripe distribution
+ * stats, stability aggregates, and a composite health score. Pi-agnostic.
+ *
+ * PREVENT-002: all queries use parameterized SQL.
+ */
+import { getStateDir } from "../../store.js";
+import { openStore } from "./utils.js";
+/** Human-readable labels for each stripe tier (0=permanent .. 4=volatile). */
+const STRIPE_LABELS = {
+    0: "Permanent",
+    1: "Epoch",
+    2: "Topic",
+    3: "Thread",
+    4: "Volatile",
+};
+/** Weight per stripe for health scoring (higher = more stable = better). */
+const STRIPE_HEALTH_WEIGHTS = {
+    0: 1.0,
+    1: 0.8,
+    2: 0.5,
+    3: 0.25,
+    4: 0.0,
+};
+/**
+ * Read cache stripe distribution from the database.
+ *
+ * Returns aggregated per-stripe buckets with stability stats and a composite
+ * health score. Returns empty buckets (all zeros) when the table is empty,
+ * so the dashboard always has a valid response.
+ */
+export function readCacheStripes(stateDir = getStateDir()) {
+    const db = openStore(stateDir);
+    // ── Fetch per-stripe aggregates ──
+    const rows = db
+        .prepare(`SELECT
+         stripe,
+         COUNT(*)                                         AS count,
+         ROUND(AVG(stability), 4)                         AS avgStability,
+         ROUND(MIN(stability), 4)                         AS minStability,
+         ROUND(MAX(stability), 4)                         AS maxStability
+       FROM cache_stripes
+       GROUP BY stripe
+       ORDER BY stripe ASC`)
+        .all();
+    // ── Latest epoch ──
+    const epochRow = db
+        .prepare(`SELECT epoch_id, MAX(assigned_at) AS last_refresh
+       FROM cache_stripes
+       WHERE epoch_id IS NOT NULL`)
+        .get();
+    // ── Total chunk count ──
+    const totalRow = db
+        .prepare("SELECT COUNT(*) AS cnt FROM cache_stripes")
+        .get();
+    const totalChunks = totalRow?.cnt ?? 0;
+    // ── Build bucket array (always 5 entries, stripe 0–4) ──
+    const rowMap = new Map();
+    for (const r of rows)
+        rowMap.set(r.stripe, r);
+    const buckets = [];
+    for (let s = 0; s <= 4; s++) {
+        const r = rowMap.get(s);
+        buckets.push({
+            stripe: s,
+            label: STRIPE_LABELS[s] ?? `Stripe ${s}`,
+            count: r?.count ?? 0,
+            avgStability: r?.avgStability ?? 0,
+            minStability: r?.minStability ?? 0,
+            maxStability: r?.maxStability ?? 0,
+        });
+    }
+    // ── Compute health score ──
+    const health = computeCacheHealth(buckets, totalChunks, stateDir);
+    return {
+        buckets,
+        health,
+        epochId: epochRow?.epoch_id ?? null,
+        lastRefreshAt: epochRow?.last_refresh ?? 0,
+        totalChunks,
+    };
+}
+/**
+ * Compute cache health from bucket distribution.
+ *
+ * Score = weighted sum of stripe fractions (normalised 0..1).
+ * Dominant tier = fraction of chunks in the top 2 stripes (0+1).
+ * Label: >= 0.8 = good, >= 0.6 = fair, >= 0.4 = degraded, else poor.
+ */
+function computeCacheHealth(buckets, total, stateDir) {
+    if (total === 0) {
+        return { score: 1, label: "good", dominantTier: 1, churnRate: 0 };
+    }
+    let weightedSum = 0;
+    let dominantCount = 0;
+    for (const b of buckets) {
+        const w = STRIPE_HEALTH_WEIGHTS[b.stripe] ?? 0;
+        weightedSum += (b.count / total) * w;
+        if (b.stripe <= 1)
+            dominantCount += b.count;
+    }
+    const dominantTier = dominantCount / total;
+    // ── Churn rate: open a new DB connection for the churn query ──
+    let churnRate = 0;
+    try {
+        const churnDb = openStore(stateDir);
+        const churnRow = churnDb
+            .prepare(`WITH
+          ranked AS (
+            SELECT chunk_id, epoch_id,
+                   ROW_NUMBER() OVER (PARTITION BY chunk_id ORDER BY assigned_at DESC) AS rn
+            FROM cache_stripes
+          ),
+          cur AS (SELECT chunk_id, epoch_id FROM ranked WHERE rn = 1),
+          prev AS (SELECT chunk_id, epoch_id FROM ranked WHERE rn = 2)
+        SELECT
+          COUNT(*) AS total_changed
+        FROM cur c
+        JOIN prev p ON c.chunk_id = p.chunk_id AND c.epoch_id != p.epoch_id`)
+            .get();
+        if (churnRow && total > 0) {
+            churnRate = Math.min(1, (churnRow.total_changed ?? 0) / total);
+        }
+    }
+    catch {
+        // churn unavailable — keep 0
+    }
+    const score = Math.round(weightedSum * 100) / 100;
+    const label = score >= 0.8 ? "good" : score >= 0.6 ? "fair" : score >= 0.4 ? "degraded" : "poor";
+    return { score, label, dominantTier, churnRate };
+}

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

@@ -0,0 +1,68 @@
+/**
+ * fts5-search.ts — FTS5 trigram search helper (L1 tier in the S44 tiered router).
+ *
+ * Runs BM25-ranked trigram search against the existing `context_chunks_trgm`
+ * virtual table (pg_trgm-equivalent, created in schema.ts). All queries are
+ * parameterized (PREVENT-002). Pi-agnostic — pure SQL over a node:sqlite
+ * DatabaseSync handle.
+ *
+ * The virtual table stores (id UNINDEXED, normalized_text) with tokenize='trigram'.
+ * FTS5 trigram queries tokenize the user query into overlapping 3-grams at query
+ * time automatically — no manual n-gram splitting needed. The BM25 rank from
+ * fts5 ranks results by trigram-overlap density.
+ */
+/**
+ * Search the context_chunks_trgm FTS5 table for checkpoints whose
+ * `normalized_text` matches `query` via trigram similarity.
+ *
+ * @param query  Raw search text (FTS5 trigram tokenizer handles n-gram splitting
+ *               automatically; input is NOT SQL-concatenated — bound as ? param).
+ * @param reader A sync-reader handle against the SQLite store.
+ * @param limit  Max results to return (default 10).
+ * @returns BM25-ranked hits, highest score first.
+ */
+export function fts5Search(query, reader, limit = 10) {
+    const sql = `
+    SELECT id, bm25(context_chunks_trgm, 0.0, 1.0) AS score
+    FROM context_chunks_trgm
+    WHERE context_chunks_trgm MATCH ?
+    ORDER BY score
+    LIMIT ?
+  `;
+    const rows = reader
+        .prepare(sql)
+        .all(query, limit);
+    return rows.map((r) => ({ id: r.id, score: r.score }));
+}
+/**
+ * Search with an optional session_id filter for scoped FTS5 recall.
+ *
+ * JOINs the FTS5 virtual table against the real `context_chunks` table on `id`
+ * so we can filter by `session_id` while still using FTS5 MATCH. The vtab
+ * stores `id UNINDEXED` (it's a row-key passthrough, not searchable text), so
+ * the join is safe and efficient via the `idx_chunks_pk` index.
+ *
+ * @param query     Raw search text.
+ * @param reader    Sync-reader handle against the SQLite store.
+ * @param sessionId Scope to a specific session (omit for cross-session search).
+ * @param limit     Max results (default 10).
+ */
+export function fts5SearchScoped(query, reader, sessionId, limit = 10) {
+    if (sessionId) {
+        const sql = `
+      SELECT t.id, bm25(context_chunks_trgm, 0.0, 1.0) AS score
+      FROM context_chunks_trgm AS t
+      INNER JOIN context_chunks AS c ON c.id = t.id
+      WHERE t.context_chunks_trgm MATCH ?
+        AND c.session_id = ?
+      ORDER BY score
+      LIMIT ?
+    `;
+        const rows = reader
+            .prepare(sql)
+            .all(query, sessionId, limit);
+        return rows.map((r) => ({ id: r.id, score: r.score }));
+    }
+    // No session filter — plain FTS5 search across all sessions.
+    return fts5Search(query, reader, limit);
+}

package/dist/src/store/sqlite/perf-samples.js

@@ -28,6 +28,7 @@ export const PERF_KINDS = [
     "cpu_sys_ms",
     "db_recompute_ms",
     "disk_write_ms",
+    "cache_health",
 ];
 function isPerfKind(k) {
     return PERF_KINDS.includes(k);

package/dist/src/store/sqlite/schema.js

@@ -1,5 +1,5 @@
 import { ACHIEVEMENT_DEFS } from "../../game/scoring.js";
-const SCHEMA_VERSION = 2;
+const SCHEMA_VERSION = 4;
 /**
  * Add `column` (with `decl`, e.g. "INTEGER") to `table` if it does not already
  * exist. Idempotent: checks PRAGMA table_info first, so it is safe to run on
@@ -353,6 +353,65 @@ export function initSchema(db) {
     );
     CREATE INDEX IF NOT EXISTS idx_perf_samples_ts ON perf_samples(ts);
     CREATE INDEX IF NOT EXISTS idx_perf_samples_kind_ts ON perf_samples(kind, ts);
+
+    -- A1 PLAN_V2 Phase 2: Message Separation tables (SCHEMA_VERSION 3).
+    -- Stable conversation thread: user + assistant turns ONLY. Tool results are
+    -- stored in tool_results to keep the conversation thread cache-predictable.
+    -- Parameterized queries only (PREVENT-002). CREATE TABLE IF NOT EXISTS is
+    -- a no-op on re-open — no migration needed.
+    CREATE TABLE IF NOT EXISTS conversation_thread (
+      id              INTEGER PRIMARY KEY AUTOINCREMENT,
+      conversation_id TEXT NOT NULL,
+      role            TEXT NOT NULL CHECK(role IN ('user','assistant')),
+      content         TEXT NOT NULL,
+      turn_index      INTEGER NOT NULL,
+      timestamp       INTEGER NOT NULL
+    );
+    CREATE INDEX IF NOT EXISTS idx_conv_thread_conv_turn
+      ON conversation_thread(conversation_id, turn_index);
+
+    -- Volatile tool results: appended at END so they never disrupt the
+    -- conversation thread prefix (the prompt-cache-friendly prefix).
+    CREATE TABLE IF NOT EXISTS tool_results (
+      id              INTEGER PRIMARY KEY AUTOINCREMENT,
+      conversation_id TEXT NOT NULL,
+      tool_call_id    TEXT NOT NULL,
+      tool_result     TEXT NOT NULL,
+      turn_index      INTEGER NOT NULL,
+      timestamp       INTEGER NOT NULL
+    );
+    CREATE INDEX IF NOT EXISTS idx_tool_results_conv_turn
+      ON tool_results(conversation_id, turn_index);
+
+    -- A2 PLAN_V2 Phase 3: Vector-Aware Cache Striping (SCHEMA_VERSION 4).
+    -- Cache stripe assignments: maps each context chunk to a cache stability
+    -- tier (stripe). Layer 0=permanent, 1=epoch, 2=topic, 3=thread, 4=volatile.
+    -- stability is the composite score (0.0-1.0) from computeStabilityScore.
+    -- Primary key is (chunk_id, epoch_id) so the same chunk can be reassigned
+    -- across epochs without conflict. See cache-stripe.ts for scoring details.
+    -- Parameterized queries only (PREVENT-002). IF NOT EXISTS is a no-op on re-open.
+    CREATE TABLE IF NOT EXISTS cache_stripes (
+      chunk_id      TEXT NOT NULL,
+      stripe        INTEGER NOT NULL,
+      stability     REAL NOT NULL,
+      assigned_at   INTEGER NOT NULL,
+      epoch_id      TEXT,
+      PRIMARY KEY (chunk_id, epoch_id)
+    );
+    CREATE INDEX IF NOT EXISTS idx_cs_stripe_stability
+      ON cache_stripes(stripe, stability DESC);
+
+    -- Embedding computation cache: avoids re-embedding content every time
+    -- stability scores are recomputed across epochs. Keyed by content_hash
+    -- (SHA-256 64-hex-char string). Embeddings are serialized via
+    -- Float64Array-to-Buffer to match the embedding_blob convention.
+    CREATE TABLE IF NOT EXISTS embedding_cache (
+      content_hash  TEXT PRIMARY KEY,
+      embedding     BLOB NOT NULL,
+      computed_at   INTEGER NOT NULL
+    );
+    CREATE INDEX IF NOT EXISTS idx_emb_cache_computed
+      ON embedding_cache(computed_at DESC);
   `);
     // Idempotent column migrations. `CREATE TABLE IF NOT EXISTS` is a no-op on a
     // pre-existing table, so new columns added to context_chunks after a store was

package/dist/src/store/sqlite.js

@@ -24,4 +24,5 @@ export * from "./sqlite/game-state.js";
 export * from "./sqlite/game-scores.js";
 export * from "./sqlite/game-achievements.js";
 export * from "./sqlite/perf-samples.js";
+export * from "./sqlite/cache-stripes.js";
 export * from "./sqlite/turns.js";

package/dist/src/store/vectorIndex.js

@@ -195,10 +195,15 @@ export async function upsertEmbedding(repoId, sessionId, checkpointId, embedding
         if (!pg)
             return;
         const lit = toVectorLiteral(embedding);
-        await pg.query(`INSERT INTO vector_index (repo_id, session_id, checkpoint_id, embedding)
-       VALUES ($1, $2, $3, $4::vector)
-       ON CONFLICT (repo_id, session_id, checkpoint_id)
-       DO UPDATE SET embedding = EXCLUDED.embedding;`, [repoId, sessionId, checkpointId, lit]);
+        const upserted = await withOpenTimeout(pg.query(`INSERT INTO vector_index (repo_id, session_id, checkpoint_id, embedding)
+         VALUES ($1, $2, $3, $4::vector)
+         ON CONFLICT (repo_id, session_id, checkpoint_id)
+         DO UPDATE SET embedding = EXCLUDED.embedding;`, [repoId, sessionId, checkpointId, lit]), (reason) => {
+            disabled = true;
+            logWarn(`upsert query timed out after ${queryTimeoutMs()}ms: ${reason}`);
+        }, queryTimeoutMs());
+        if (upserted === undefined)
+            return; // timed out — non-fatal
     }
     catch (err) {
         disabled = true;
@@ -227,7 +232,12 @@ export async function searchAsync(query, opts = {}) {
             params.push(repoId);
         }
         sql += " ORDER BY embedding <=> $1::vector LIMIT $2";
-        const res = await pg.query(sql, params);
+        const res = await withOpenTimeout(pg.query(sql, params), (reason) => {
+            disabled = true;
+            logWarn(`search query timed out after ${queryTimeoutMs()}ms: ${reason}`);
+        }, queryTimeoutMs());
+        if (!res)
+            return []; // query timed out — degrade to empty (non-fatal)
         return res.rows.map((r) => ({
             repoId: r.repo_id,
             sessionId: r.session_id,
@@ -252,6 +262,20 @@ function closeTimeoutMs() {
     }
     return DEFAULT_CLOSE_TIMEOUT_MS;
 }
+/** Default cap for a PGlite search query. A WASM query that never resolves would
+ *  otherwise hang the calling turn (and node --test) forever — the catch below
+ *  only handles thrown errors, not a stalled promise. 0 disables the guard
+ *  (preserve original unbounded behavior). Mirrors the open/close timeouts. */
+export const DEFAULT_QUERY_TIMEOUT_MS = 10_000;
+function queryTimeoutMs() {
+    const raw = process.env.MEGACOMPACT_PGLITE_QUERY_TIMEOUT_MS;
+    if (raw !== undefined) {
+        const n = Number(raw);
+        if (Number.isFinite(n) && n >= 0)
+            return n;
+    }
+    return DEFAULT_QUERY_TIMEOUT_MS;
+}
 /**
  * Race a promise against a timeout. If the inner promise never settles, the
  * timer fires, resolves with `undefined`, and the caller proceeds. This mirrors

package/dist/src/tiered-router/types.js

@@ -0,0 +1,18 @@
+/**
+ * tiered-router/types.ts — S44 tiered router types, interfaces, and defaults.
+ *
+ * Shell file: pure data/type definitions. Zero runtime dependencies on store
+ * internals. Imported by tieredRouter.ts and any consumer that needs the
+ * RecallResult / TieredRouterMetrics shapes.
+ */
+// ---------------------------------------------------------------------------
+// Defaults
+// ---------------------------------------------------------------------------
+/** Default LRU cache size. */
+export const DEFAULT_CACHE_SIZE = 256;
+/** Default latency budgets per tier (ms). */
+export const DEFAULT_BUDGET_L0_MS = 1;
+export const DEFAULT_BUDGET_L1_MS = 50;
+export const DEFAULT_BUDGET_L2_MS = 500;
+/** Default cadence for tiered-metrics logging (every N queries). */
+export const DEFAULT_LOG_CADENCE = 100;