diff --git a/docs/benchmarks/index-drift-cg33.md b/docs/benchmarks/index-drift-cg33.md
new file mode 100644
index 0000000..548e793
--- /dev/null
+++ b/docs/benchmarks/index-drift-cg33.md
@@ -0,0 +1,103 @@
+# Index drift: incremental sync vs. full rebuild (CG-33)
+
+Measured 2026-08-06. A live, auto-sync-maintained index **does not converge** to
+a clean full rebuild of the identical working tree. On codegraph's own repo,
+**4.3% of distinct edges were wrong**, in both directions, overwhelmingly
+`calls` edges.
+
+This matters because it is silent: nothing warns, nothing surfaces it, and the
+README tells users the index is never stale and there is nothing to re-run.
+Retrieval quality decays invisibly, and the user-visible symptom — an agent
+falling back to Read — reads as "codegraph isn't very good" rather than "this
+index needs rebuilding."
+
+## Result
+
+Subject: codegraph's own `.codegraph/codegraph.db`, long-lived and
+incrementally synced, against a full rebuild of the same tree with the same
+build. Edges compared as distinct `(source, target, kind)` triples.
+
+| | count |
+|---|---|
+| distinct edge triples (rebuild) | 28,809 |
+| in rebuild but **missing** from live | **751** |
+| in live but **absent** from rebuild (stale) | **476** |
+| **total divergent** | **1,227 — 4.3%** |
+
+Missing edges by kind: `calls=635`, `contains=38`, `references=34`,
+`instantiates=21`, `imports=13`, `extends=10`.
+
+### Raw counts hide it
+
+Raw edge **rows** were 39,845 live vs 40,122 rebuilt — a benign-looking +0.7%.
+The divergence is bidirectional, so a net-count check nets it out and reports
+almost nothing wrong. **Any drift detector must compare edge sets, not totals.**
+
+### The indexer is deterministic
+
+Control, rebuild vs rebuild on the same tree and build: **0 differing edges**
+(28,809 both runs). So the live-vs-rebuild delta is not run-to-run noise.
+
+### It is resolution, not residue
+
+Node sets are identical — `files` 501 = 501, `nodes` 10,110 = 10,110,
+heuristic edges 36 = 36 — and every integrity check is 0 on *both* indexes:
+no duplicate nodes, no orphan edges, no nodes referencing a missing file row.
+
+Nothing accumulates. Cross-file **resolution** goes stale.
+
+## Likely mechanism
+
+`ReferenceResolver` resolves calls and imports by name-matching and the import
+graph across the **whole** project. Incremental sync re-parses and re-resolves
+only the changed file, so:
+
+- edges from *other* files into changed symbols are never recomputed → stale
+  edges retained (the 476);
+- edges that should newly form from unchanged files into changed symbols are
+  never created → missing edges (the 751).
+
+Start in `src/sync/` and `src/resolution/` — specifically what scope is
+re-resolved on a single-file change.
+
+## Why it degrades retrieval
+
+Graph mass (RWR) is **relative and normalized**, so call edges missing elsewhere
+inflate an unaffected file's share of the mass. Explore ranks files by that mass
+(`allocateExploreBudget` weights on it), so drift silently promotes files that
+should rank low.
+
+Observed on a private application repo under heavy development: a generated
+ambient-types file carried graph mass **0.24750** on the drifted index vs
+**0.13119** on a clean rebuild (~1.9×), and score **49.0** vs **27.0**. On the
+drifted index it took **60.7%** of an explore envelope and starved the file the
+agent had actually named by symbol, which rendered **251 chars of a 10,970
+reservation**. After a full re-index — no code change — the same query answers
+correctly. That incident is what prompted this measurement; see CG-24.
+
+Severity scales with churn and index age. codegraph's own repo shows 4.3%;
+a repo under heavier active development plausibly drifts further.
+
+## Reproducing
+
+`scripts/agent-eval/diff-index-drift.mjs` is read-only and diffs two indexes.
+Snapshot the live index **before** rebuilding — the original artifact for this
+investigation was destroyed by re-indexing over it:
+
+```bash
+cp .codegraph/codegraph.db /tmp/live.db          # snapshot FIRST
+node dist/bin/codegraph.js index .               # full rebuild
+node scripts/agent-eval/diff-index-drift.mjs /tmp/live.db .codegraph/codegraph.db
+```
+
+Exit code is 0 when converged, 1 when drifted. To re-confirm determinism, diff
+two consecutive rebuilds — that must report 0.
+
+## Note on probing an index
+
+The index file is `.codegraph/codegraph.db`. There is no `graph.db`. `sqlite3`
+against a mistyped path **creates an empty database** rather than failing, and
+every subsequent query then answers from an empty schema — which reads exactly
+like a stale pre-migration index. That produced a wrong root cause during this
+investigation. `diff-index-drift.mjs` checks `existsSync` before opening for
+exactly this reason.
diff --git a/scripts/agent-eval/diff-index-drift.mjs b/scripts/agent-eval/diff-index-drift.mjs
new file mode 100644
index 0000000..9459549
--- /dev/null
+++ b/scripts/agent-eval/diff-index-drift.mjs
@@ -0,0 +1,102 @@
+#!/usr/bin/env node
+/**
+ * Diff two CodeGraph indexes of the SAME tree — typically a live,
+ * incrementally-synced `.codegraph/codegraph.db` against a clean full rebuild
+ * of the identical working tree (CG-33).
+ *
+ * Non-destructive: it only reads. Rebuilding is the caller's job, so the live
+ * index is never clobbered by the tool measuring it — the mistake that cost the
+ * original CG-33 artifact.
+ *
+ *   # snapshot the live index BEFORE touching it
+ *   cp .codegraph/codegraph.db /tmp/live.db
+ *   node dist/bin/codegraph.js index .
+ *   node scripts/agent-eval/diff-index-drift.mjs /tmp/live.db .codegraph/codegraph.db
+ *
+ * Edges are compared as distinct `(source, target, kind)` triples. Raw row
+ * counts are NOT a drift signal: a bidirectional divergence nets out. On the
+ * codegraph repo the raw counts differed by +0.7% while 4.3% of distinct edges
+ * were actually wrong.
+ */
+import { DatabaseSync } from 'node:sqlite';
+import { existsSync } from 'node:fs';
+
+const [livePath, rebuiltPath] = process.argv.slice(2);
+if (!livePath || !rebuiltPath) {
+  console.error('usage: diff-index-drift.mjs <live.db> <rebuilt.db>');
+  process.exit(2);
+}
+for (const p of [livePath, rebuiltPath]) {
+  if (!existsSync(p)) {
+    // node:sqlite CREATES a missing file rather than failing, which silently
+    // yields an empty schema and a confident, wrong conclusion. Refuse first.
+    console.error(`not found: ${p}`);
+    process.exit(2);
+  }
+}
+
+const open = (p) => new DatabaseSync(p, { readOnly: true });
+const live = open(livePath);
+const rebuilt = open(rebuiltPath);
+
+const scalar = (db, q) => db.prepare(q).get().n;
+const edgeKey = (r) => `${r.source}\u0000${r.target}\u0000${r.kind}`;
+
+const liveEdges = live.prepare('select source, target, kind from edges').all();
+const rebuiltEdges = rebuilt.prepare('select source, target, kind from edges').all();
+const liveSet = new Set(liveEdges.map(edgeKey));
+const rebuiltSet = new Set(rebuiltEdges.map(edgeKey));
+
+const missing = rebuiltEdges.filter((r) => !liveSet.has(edgeKey(r)));   // should exist, doesn't
+const stale = liveEdges.filter((r) => !rebuiltSet.has(edgeKey(r)));      // exists, shouldn't
+const divergent = missing.length + stale.length;
+
+const byKind = (rows) => {
+  const m = new Map();
+  for (const r of rows) m.set(r.kind, (m.get(r.kind) ?? 0) + 1);
+  return [...m].sort((a, b) => b[1] - a[1]).map(([k, v]) => `${k}=${v}`).join(', ') || '(none)';
+};
+
+const pct = (n, d) => (d ? ((n / d) * 100).toFixed(1) : '0.0');
+
+console.log(`live    ${livePath}`);
+console.log(`rebuilt ${rebuiltPath}`);
+console.log('');
+console.log('counts                 live      rebuilt');
+for (const [label, q] of [
+  ['files', 'select count(*) n from files'],
+  ['nodes', 'select count(*) n from nodes'],
+  ['edges (rows)', 'select count(*) n from edges'],
+  // Grouped rather than `count(distinct a || b || c)`: bare concatenation has no
+  // separator, so `(ab, c)` and `(a, bc)` would collapse into one.
+  ['edges (distinct)', 'select count(*) n from (select distinct source, target, kind from edges)'],
+  ['heuristic edges', "select count(*) n from edges where provenance='heuristic'"],
+]) {
+  console.log(`  ${label.padEnd(20)} ${String(scalar(live, q)).padEnd(9)} ${scalar(rebuilt, q)}`);
+}
+
+console.log('');
+console.log('edge divergence (distinct triples)');
+console.log(`  missing from live: ${missing.length}  — ${byKind(missing)}`);
+console.log(`  stale in live:     ${stale.length}  — ${byKind(stale)}`);
+console.log(`  TOTAL divergent:   ${divergent}  (${pct(divergent, rebuiltSet.size)}% of ${rebuiltSet.size})`);
+
+// Integrity checks — these separate "resolution went stale" (edges wrong, nodes
+// identical) from "residue accumulated" (duplicate/orphan rows). CG-33 is the
+// former: on the codegraph repo every check below was 0 on BOTH indexes.
+console.log('');
+console.log('integrity                        live   rebuilt');
+for (const [label, q] of [
+  ['duplicate nodes', 'select count(*) n from (select file_path,name,kind,start_line from nodes group by 1,2,3,4 having count(*)>1)'],
+  ['orphan edges', 'select count(*) n from edges e where not exists(select 1 from nodes where id=e.source) or not exists(select 1 from nodes where id=e.target)'],
+  ['nodes w/ missing file row', 'select count(*) n from nodes nd where not exists(select 1 from files f where f.path=nd.file_path)'],
+]) {
+  console.log(`  ${label.padEnd(30)} ${String(scalar(live, q)).padEnd(6)} ${scalar(rebuilt, q)}`);
+}
+
+console.log('');
+console.log(divergent === 0
+  ? 'CONVERGED — the synced index matches a full rebuild.'
+  : `DRIFTED — ${divergent} edges differ. Rebuild-vs-rebuild is 0 (the indexer is deterministic), so this is sync divergence, not noise.`);
+
+process.exitCode = divergent === 0 ? 0 : 1;