docs: publish five-language Compass vs Graphify evaluation

PERFORMANCE.md

@@ -8,6 +8,13 @@ The reproducible real-repository harness, operator commands, correctness gates,
 and optional explicit Graphify comparison are documented in
 [`benchmarks/performance/README.md`](benchmarks/performance/README.md).
 
+The [2026-10-02 real-repository comparison](benchmarks/agent_query/reports/2026-10-02-main-vs-graphify/REPORT.md)
+retains indexing, query, CPU and memory observations for Compass 0.4.1 main
+and Graphify 0.9.74 across five languages. It discloses the busy shared host,
+different extraction scopes, storage/clustering configurations and every
+failed observation. This diagnostic capture does not establish a controlled
+performance ranking or replace the qualification baseline below.
+
 ## Baseline policy
 
 A benchmark records:

benchmarks/agent_query/README.md

@@ -9,6 +9,14 @@ question/evidence matrix, including ask, communities, clusters, and god nodes.
 The audit report distinguishes completed checks from surfaces still awaiting
 source or design-quality judgments.
 
+The [2026-10-02 comparison](reports/2026-10-02-main-vs-graphify/REPORT.md)
+records Compass main `b7b7fa81` (0.4.1) against Graphify 0.9.74 on these five
+language panels, with JSON/SQLite and base/native-Leiden configurations kept
+separate. It retains 120 builds, 1,020 query workflows, source oracles, raw
+transcripts and path-normalized method snapshots. These selected development
+inputs and shared-host timings are diagnostic evidence; text recall and the
+post-capture directional audit have distinct scoring requirements.
+
 Five suites share the harness:
 
 | Suite | Questions | Shape |

benchmarks/agent_query/reports/2026-10-02-main-vs-graphify/data/ask-registration.json

@@ -0,0 +1,7 @@
+{
+  "protocol": "Ten existing source-reviewed caller/callee tasks, identical question text and 2000-token budget; Compass ask versus Graphify query. Three repetitions, all four variants. These reuse source facts from the structured panel; do not count them as independent source labels.",
+  "registeredAt": "2026-10-02T18:10:13.794695+00:00",
+  "registeredBeforeAnyQueryExecution": true,
+  "suite": "${COMPASS_CHECKOUT}/benchmarks/agent_query/suite_ask.toml",
+  "suiteSha256": "ea1306f24b1f612f78f5d4e579df21d07ac8537dc49bbb403be1eddce71110e4"
+}

benchmarks/agent_query/reports/2026-10-02-main-vs-graphify/data/census-scorer-amendment.json

@@ -0,0 +1,6 @@
+{
+  "afterInitialGraphCaptureBeforeCensusScoring": true,
+  "at": "2026-10-02T18:12:32.505321+00:00",
+  "censusScriptSha256": "17beeed9c6932c94146462312cfcf24c1d27a637c4f23e4094e62444a5c96086",
+  "reason": "Graphify clustered AST nodes omit kind/type in their published schema, including normal function nodes. The scorer must not reject a field solely for an absent optional kind. Require the same exact source file, identifier line, name, unique declaration, and direct containment by the exact source type. Allow absent kind or field/property/attribute; reject contradictory explicit kinds. Source files and expected declarations are unchanged."
+}

benchmarks/agent_query/reports/2026-10-02-main-vs-graphify/data/completion-verification.json

@@ -0,0 +1,20 @@
+{
+  "buildObservations": 120,
+  "compassHead": "b7b7fa819b127464566430cc454fd2085d5d5b62",
+  "queryObservations": 1020,
+  "queryTimeouts": 1,
+  "scriptsSha256": {
+    "census.py": "17beeed9c6932c94146462312cfcf24c1d27a637c4f23e4094e62444a5c96086",
+    "directional_audit.py": "cdb4d0785bbd9f9e5ab9ffae332e65e38dcb4acff07aa05c21b02524e1d5d5fb",
+    "evaluate.py": "e09529a5ccdbec34b87d1fe64207c5ad6eb6b3c78304c36dee5af4077dcd2aff",
+    "source_audit.py": "4397599684a69e133b65b25846f53c9f1b3c7f750a4e26edc3187c953f6870a3",
+    "summarize.py": "7c83ecb22f3f8efd066bc543f8a9b43cd5736f53b8927ab42a1f80e10047368b",
+    "supplement.py": "0d4dd7f466b2096cf6327770e8ac7551af37fa475260e9bb01e7ee6bac4b7c88",
+    "transitions.py": "f0fef0559231be6236cba9590374501108bdc4d33bfa3aaf128661aab0948bcd",
+    "verify_evidence.py": "9af80d5dd34f6487ee6c180e3ca239346efec9d78bb80bf1861257dab4b2851e"
+  },
+  "sourceVerification": "All five pinned clean corpus HEADs, suite source digests and before-build call-oracle source digests passed.",
+  "verification": "Graph digests, query graph identities, binary identities, both Graphify package source trees and benchmark helper identities unchanged.",
+  "verifiedAt": "2026-10-02T19:07:08.029053+00:00",
+  "verifiedFrozenGraphs": 120
+}