coderadar-rs 0.9.3__tar.gz → 0.9.4__tar.gz

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/Cargo.lock

@@ -445,7 +445,7 @@ checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
 
 [[package]]
 name = "core_indexer"
-version = "0.9.3"
+version = "0.9.4"
 dependencies = [
  "crossbeam-channel",
  "dashmap",
@@ -1640,9 +1640,9 @@ dependencies = [
 
 [[package]]
 name = "macrame-db"
-version = "0.17.0"
+version = "0.18.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e2a46b8789555136ad929e72df4ebb5b7cda21e8c6512dc45a1dec54ad2a54d"
+checksum = "6c07608eae01b1acf1fa71f739e92e9cc4c5ab90c8ebadc82b36eee500fd8a1d"
 dependencies = [
  "bincode",
  "libsql",

coderadar_rs-0.9.4/Cargo.toml

@@ -0,0 +1,10 @@
+[workspace]
+members = ["core_indexer"]
+resolver = "2"
+
+[workspace.package]
+version = "0.9.4"
+edition = "2021"
+license = "MIT"
+homepage = "https://github.com/opticsWolf/coderadar"
+repository = "https://github.com/opticsWolf/coderadar"

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coderadar-rs
-Version: 0.9.3
+Version: 0.9.4
 Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Developers
 Classifier: License :: OSI Approved :: MIT License
@@ -27,12 +27,15 @@ Provides-Extra: dev
 Provides-Extra: visualizers
 License-File: LICENSE
 Summary: Live semantic graph of your codebase — incremental, queryable, LLM-writable
+Home-Page: https://github.com/opticsWolf/coderadar
 Author: CodeRadar Team
 License: MIT
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
+Project-URL: Homepage, https://github.com/opticsWolf/coderadar
+Project-URL: Repository, https://github.com/opticsWolf/coderadar
 
-# CodeRadar v0.9.2
+# CodeRadar v0.9.4
 
 [![CI](https://github.com/opticsWolf/coderadar/actions/workflows/ci.yml/badge.svg)](https://github.com/opticsWolf/coderadar/actions/workflows/ci.yml)
 [![PyPI](https://img.shields.io/pypi/v/coderadar-rs?label=pypi)](https://pypi.org/project/coderadar-rs/)
@@ -41,6 +44,8 @@ Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
 [![License](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
 [![Languages](https://img.shields.io/badge/languages-41-brightgreen)]()
 
+[Homepage](https://github.com/opticsWolf/coderadar) · [Repository](https://github.com/opticsWolf/coderadar)
+
 **Live semantic graph of your codebase — incremental, queryable, LLM-writable.**
 
 CodeRadar maintains an incrementally updatable graph of your code's logical structure, enabling LLMs and developer tools to both **query** and **safely rewrite** code through a unified pipeline.
@@ -89,7 +94,7 @@ Rust Core (ProjectedGraph, Tree-sitter 41-lang, Parallel Extraction,
 | Metric | Value |
 |--------|-------|
 | **Languages indexed** | 41 (12 Tier 1, 29 Tier 2, 330+ Tier 3) |
-| **Tests** | 1128 (369 Rust + 759 Python) |
+| **Tests** | 1139 passing (372 Rust + 767 Python; 1 Python skipped) |
 | **MCP Tools** | 22 — 17 `codegraph_*` (explore, search, node, affected, query, search_similar, compute_embeddings, module_children, as_of, traverse, get_smells, dead_code, find_clones, find_scaffolding, reindex, update_file, set_project) + 5 `coderadar_*` (resolve, replace_body, update_signature, rename, create_entity) |
 | **Query surface** | Pest structural + Macrame agent traversals + vector search |
 | **Frameworks** | Django, Flask, FastAPI, Go, Actix, Express, Spring Boot, Laravel, ASP.NET, Rails, NestJS, Vue Router, React Router |
@@ -335,13 +340,17 @@ cold-start design in [`docs/v0.8-p1-cold-start-design.md`](docs/v0.8-p1-cold-sta
 - **Shell-friendly entity IDs (E7).** Stored IDs keep OS-native separators (FK stability);
   display renders forward slashes with the redundant `./` dropped, and every pasted variant
   resolves back to the stored key.
-- **macrame-db 0.17 (schema v19, auto-rungs v15→v19 on open).** Builders for the now-`#[non_exhaustive]`
-  `ConceptUpsert`, `Vec<EdgeBelief>` replay folds, distinct `BulkInterrupted`
-  reporting, and `Database::analyze()` planner statistics after the bulk flush.
-  Two perf fixes fell out: the ledger-revision stamp reads `MAX(seq_id)` instead
-  of a full `reconstruct` fold (1.7 s → 0.08 s per analyze on a large log), and
-  content-hash-gated upserts make no-op analyzes write zero rows (see
-  [`docs/macrame-0.15-upgrade-notes.md`](docs/macrame-0.15-upgrade-notes.md)).
+- **macrame-db 0.18 (schema v21, auto-rungs v15→v21 on open).** CodeRadar stores
+  compact namespaced metadata (`kind`, `file_path`, `content_hash`) in
+  `concepts.extra`, indexes the file path for precise stale-row cleanup, and
+  gives framework edges stable namespaced kinds such as
+  `synthetic:django:handles` instead of flattening their types. This is a
+  forward-only database upgrade: a 0.17 binary cannot open v21 or fold a
+  concept written by 0.18. Existing 0.17 concept content remains canonical and
+  is rewritten with `extra` on the next analyze. Existing no-op-write filtering
+  and the `MAX(seq_id)` ledger-revision stamp remain in place; see the
+  [integration notes](docs/macrame-0.18-integration.md) for the adopted fields
+  and the KV deferral.
 
 ## v0.8.1–v0.8.14 Highlights — the dogfood batch
 

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/README.md

@@ -1,4 +1,4 @@
-# CodeRadar v0.9.2
+# CodeRadar v0.9.4
 
 [![CI](https://github.com/opticsWolf/coderadar/actions/workflows/ci.yml/badge.svg)](https://github.com/opticsWolf/coderadar/actions/workflows/ci.yml)
 [![PyPI](https://img.shields.io/pypi/v/coderadar-rs?label=pypi)](https://pypi.org/project/coderadar-rs/)
@@ -7,6 +7,8 @@
 [![License](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
 [![Languages](https://img.shields.io/badge/languages-41-brightgreen)]()
 
+[Homepage](https://github.com/opticsWolf/coderadar) · [Repository](https://github.com/opticsWolf/coderadar)
+
 **Live semantic graph of your codebase — incremental, queryable, LLM-writable.**
 
 CodeRadar maintains an incrementally updatable graph of your code's logical structure, enabling LLMs and developer tools to both **query** and **safely rewrite** code through a unified pipeline.
@@ -55,7 +57,7 @@ Rust Core (ProjectedGraph, Tree-sitter 41-lang, Parallel Extraction,
 | Metric | Value |
 |--------|-------|
 | **Languages indexed** | 41 (12 Tier 1, 29 Tier 2, 330+ Tier 3) |
-| **Tests** | 1128 (369 Rust + 759 Python) |
+| **Tests** | 1139 passing (372 Rust + 767 Python; 1 Python skipped) |
 | **MCP Tools** | 22 — 17 `codegraph_*` (explore, search, node, affected, query, search_similar, compute_embeddings, module_children, as_of, traverse, get_smells, dead_code, find_clones, find_scaffolding, reindex, update_file, set_project) + 5 `coderadar_*` (resolve, replace_body, update_signature, rename, create_entity) |
 | **Query surface** | Pest structural + Macrame agent traversals + vector search |
 | **Frameworks** | Django, Flask, FastAPI, Go, Actix, Express, Spring Boot, Laravel, ASP.NET, Rails, NestJS, Vue Router, React Router |
@@ -301,13 +303,17 @@ cold-start design in [`docs/v0.8-p1-cold-start-design.md`](docs/v0.8-p1-cold-sta
 - **Shell-friendly entity IDs (E7).** Stored IDs keep OS-native separators (FK stability);
   display renders forward slashes with the redundant `./` dropped, and every pasted variant
   resolves back to the stored key.
-- **macrame-db 0.17 (schema v19, auto-rungs v15→v19 on open).** Builders for the now-`#[non_exhaustive]`
-  `ConceptUpsert`, `Vec<EdgeBelief>` replay folds, distinct `BulkInterrupted`
-  reporting, and `Database::analyze()` planner statistics after the bulk flush.
-  Two perf fixes fell out: the ledger-revision stamp reads `MAX(seq_id)` instead
-  of a full `reconstruct` fold (1.7 s → 0.08 s per analyze on a large log), and
-  content-hash-gated upserts make no-op analyzes write zero rows (see
-  [`docs/macrame-0.15-upgrade-notes.md`](docs/macrame-0.15-upgrade-notes.md)).
+- **macrame-db 0.18 (schema v21, auto-rungs v15→v21 on open).** CodeRadar stores
+  compact namespaced metadata (`kind`, `file_path`, `content_hash`) in
+  `concepts.extra`, indexes the file path for precise stale-row cleanup, and
+  gives framework edges stable namespaced kinds such as
+  `synthetic:django:handles` instead of flattening their types. This is a
+  forward-only database upgrade: a 0.17 binary cannot open v21 or fold a
+  concept written by 0.18. Existing 0.17 concept content remains canonical and
+  is rewritten with `extra` on the next analyze. Existing no-op-write filtering
+  and the `MAX(seq_id)` ledger-revision stamp remain in place; see the
+  [integration notes](docs/macrame-0.18-integration.md) for the adopted fields
+  and the KV deferral.
 
 ## v0.8.1–v0.8.14 Highlights — the dogfood batch
 

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/core_indexer/Cargo.toml

@@ -3,6 +3,8 @@ name = "core_indexer"
 version.workspace = true
 edition.workspace = true
 license.workspace = true
+homepage.workspace = true
+repository.workspace = true
 
 [lib]
 name = "core_indexer"
@@ -23,7 +25,7 @@ xxhash-rust = { version = "0.8", features = ["xxh3"] }
 git2 = { version = "0.21" }
 zstd = "0.13"
 # slotmap and arc-swap removed in v3.5 — replaced by HashMap<String, Arc<T>> + RwLock
-macrame-db = "0.17"
+macrame-db = "0.18"
 tokio = { version = "1", features = ["rt-multi-thread"] }
 pest = "2.7"
 pest_derive = "2.7"

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/core_indexer/src/graph/cold_start.rs

@@ -208,7 +208,7 @@ fn restore_synthetic_edges(
     state: &MaterializedState,
     stats: &mut ColdStartStats,
 ) {
-    // Defensive (source, target, kind) triple dedup. macrame 0.15 enforces
+    // Defensive (source, target, kind) triple dedup. macrame 0.15+ enforces
     // a single open interval per triple, so duplicates are impossible in
     // practice; the guard stays cheap in case legacy rows ever appear.
     let mut seen: HashSet<(String, String, String)> = HashSet::new();
@@ -428,12 +428,12 @@ mod tests {
         for c in &concepts {
             map.insert(
                 c.id.clone(),
-                // macrame 0.17: NodeAttributes is #[non_exhaustive]
+                // macrame 0.17+: NodeAttributes is #[non_exhaustive]
                 // (embedding_model defaults None, as before).
                 NodeAttributes::new(c.id.clone(), c.title.clone(), c.content.clone()),
             );
         }
-        // macrame 0.17: MaterializedState is #[non_exhaustive] —
+        // macrame 0.17+: MaterializedState is #[non_exhaustive] —
         // assemble from empty(); fields stay pub and assignable.
         let mut state = MaterializedState::empty("2026-02-08T00:00:00.000000Z");
         state.seq_anchor = 42;

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/core_indexer/src/graph/persistence.rs

@@ -316,19 +316,11 @@ impl CodeGraph {
         // Persist to Macrame if store attached. Re-registering a synthetic
         // edge that is already open is skipped (same fact), otherwise the
         // re-run would abort on the single-open guard and the error would be
-        // silently dropped below. Two boundary guards, both enforced here so
-        // no caller can trip them:
-        //
-        //  * macrame's edge-type validator accepts only `[A-Z0-9]+`, so an
-        //    underscore kind (`DEPENDS_ON`) would fail the whole
-        //    all-or-nothing batch and lose every edge in it. The underscore
-        //    is stripped at this boundary; the in-memory indices are
-        //    kind-agnostic, so nothing else changes.
-        //  * `links` has FKs to `concepts` on both endpoints, so edges whose
-        //    endpoints are not persisted concepts (e.g. `django:route:...`
-        //    route nodes) cannot be asserted; the batch drops the failing
-        //    write rather than failing the registration (best-effort, as
-        //    before).
+        // silently dropped below. Macrame 0.18 accepts namespaced kinds, so
+        // preserve the caller's semantic edge type rather than flattening it.
+        // `links` still has FKs to `concepts` on both endpoints, so edges whose
+        // endpoints are not persisted concepts (e.g. `django:route:...` route
+        // nodes) cannot be asserted; the batch drops those best-effort.
         if let Some(store) = self.store.as_ref() {
             let ts_now = crate::storage::now_iso8601();
             // Unreadable ledger → assert everything (the pre-existing
@@ -341,13 +333,12 @@ impl CodeGraph {
                 .collect();
             let batch: Vec<_> = edges
                 .iter()
-                .map(|(source_id, target_id, kind)| {
-                    let kind = kind.replace('_', "");
-                    (source_id.clone(), target_id.clone(), kind)
-                })
                 .filter(|(source_id, target_id, kind)| {
                     !open.contains(&edge_key(source_id, target_id, kind))
                 })
+                .map(|(source_id, target_id, kind)| {
+                    (source_id.clone(), target_id.clone(), kind.clone())
+                })
                 .map(|(source_id, target_id, kind)| {
                     macrame::graph::EdgeAssertion::new(
                         source_id.as_str(),

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/core_indexer/src/graph/tests/persistence_tests.rs

@@ -528,10 +528,11 @@ fn synthetic_pairs_not_persisted_as_calls_and_stale_edges_retracted() {
             .map(|(_, _, k)| k)
             .collect()
     };
-    // Register strips the underscore (macrame kinds are [A-Z0-9]+).
+    // Macrame 0.18 preserves the caller's edge-kind spelling, including
+    // underscores, instead of collapsing distinct kinds at this boundary.
     assert_eq!(
         kinds_for(&run_id, &combine_id),
-        vec!["DEPENDSON".to_string()]
+        vec!["DEPENDS_ON".to_string()]
     );
     assert_eq!(kinds_for(&main_id, &run_id), vec!["CALLS".to_string()]);
 
@@ -556,7 +557,26 @@ fn synthetic_pairs_not_persisted_as_calls_and_stale_edges_retracted() {
     assert_eq!(retired, 1, "only the legacy synthetic-as-CALLS row closes");
     assert_eq!(
         kinds_for(&run_id, &combine_id),
-        vec!["DEPENDSON".to_string()]
+        vec!["DEPENDS_ON".to_string()]
     );
     assert_eq!(kinds_for(&main_id, &run_id), vec!["CALLS".to_string()]);
 }
+
+#[test]
+fn namespaced_synthetic_edge_kind_is_persisted_verbatim() {
+    let (graph, _dir) = graph_with_temp_store();
+    index_source(&graph, "def handler(): pass\n", "views.py");
+    let source = "views.py::module";
+    let target = "views.py::handler";
+    graph
+        .register_synthetic_edge(source, target, "synthetic:django:handles")
+        .expect("register namespaced kind");
+
+    let store = graph.store.as_ref().unwrap();
+    let open = store
+        .open_edge_triples(&crate::storage::now_iso8601())
+        .expect("read open edges");
+    assert!(open
+        .iter()
+        .any(|(s, t, kind)| { s == source && t == target && kind == "synthetic:django:handles" }));
+}

{coderadar_rs-0.9.3 → coderadar_rs-0.9.4}/core_indexer/src/lib.rs

@@ -845,7 +845,7 @@ fn analyze(
 }
 
 fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> AnalyzeOutcome {
-    use crate::types::Language;
+    use crate::types::{Language, ParseQuality};
     use std::fs;
 
     let config = active_config();
@@ -901,6 +901,10 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
     // reported plain success.
     let mut failures: Vec<String> = Vec::new();
     let mut panicked_workers: usize = 0;
+    // Reconcile absent ledger concepts only after a complete, clean scan. A
+    // walk, file-read, extraction, worker, or recovered-parse failure must
+    // never make good concepts look stale merely because source was missed.
+    let mut reconciliation_safe = root_path.is_dir();
     let mut all_concepts: Vec<macrame::ConceptUpsert> = Vec::new();
 
     if root_path.is_dir() {
@@ -915,7 +919,14 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
         for entry in project_walk(root, &config.project, extra_excludes) {
             match entry {
                 Ok(entry) => {
-                    if !entry.file_type().map_or(false, |ft| ft.is_file()) {
+                    let file_type = match entry.file_type() {
+                        Some(file_type) => file_type,
+                        None => {
+                            reconciliation_safe = false;
+                            continue;
+                        }
+                    };
+                    if !file_type.is_file() {
                         continue;
                     }
                     let path = entry.path();
@@ -931,26 +942,11 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
                         if CodeGraph::ts_language(&language).is_none() {
                             continue;
                         }
-                        if let Ok(source) = fs::read_to_string(path) {
-                            // F14: mint the canonical id form at write time —
-                            // the walk's spelling embeds the root used (`.`
-                            // vs absolute), so it is normalized here.
-                            let canon = crate::graph::module_resolution::canonical_file_form(
-                                &path.to_string_lossy(),
-                            );
-                            tasks.push(FileTask {
-                                path: canon,
-                                source,
-                                language,
-                            });
-                        }
-                    } else {
-                        // Files without extension (Dockerfile, CMakeLists.txt)
-                        let language = Language::from_filename(&path.to_string_lossy());
-                        if language != Language::OtherTen
-                            && CodeGraph::ts_language(&language).is_some()
-                        {
-                            if let Ok(source) = fs::read_to_string(path) {
+                        match fs::read_to_string(path) {
+                            Ok(source) => {
+                                // F14: mint the canonical id form at write time —
+                                // the walk's spelling embeds the root used (`.`
+                                // vs absolute), so it is normalized here.
                                 let canon = crate::graph::module_resolution::canonical_file_form(
                                     &path.to_string_lossy(),
                                 );
@@ -960,10 +956,41 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
                                     language,
                                 });
                             }
+                            Err(e) => {
+                                reconciliation_safe = false;
+                                failures.push(format!("{}: {e}", path.display()));
+                            }
+                        }
+                    } else {
+                        // Files without extension (Dockerfile, CMakeLists.txt)
+                        let language = Language::from_filename(&path.to_string_lossy());
+                        if language != Language::OtherTen
+                            && CodeGraph::ts_language(&language).is_some()
+                        {
+                            match fs::read_to_string(path) {
+                                Ok(source) => {
+                                    let canon =
+                                        crate::graph::module_resolution::canonical_file_form(
+                                            &path.to_string_lossy(),
+                                        );
+                                    tasks.push(FileTask {
+                                        path: canon,
+                                        source,
+                                        language,
+                                    });
+                                }
+                                Err(e) => {
+                                    reconciliation_safe = false;
+                                    failures.push(format!("{}: {e}", path.display()));
+                                }
+                            }
                         }
                     }
                 }
-                Err(_) => {} // permission errors etc.
+                Err(e) => {
+                    reconciliation_safe = false;
+                    failures.push(format!("project walk: {e}"));
+                }
             }
         }
 
@@ -1052,6 +1079,18 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
                 let mut proj = (*graph.snapshot()).clone();
                 for chunk_results in all_results {
                     for (fragment, concepts) in chunk_results {
+                        // Tree-sitter can recover a partial projection from
+                        // malformed source without returning an extraction
+                        // error. Keep old ledger concepts until every file is
+                        // cleanly parsed; absence from a recovered AST is not
+                        // reliable evidence of deletion.
+                        if fragment
+                            .modules
+                            .values()
+                            .any(|module| module.parse_quality != ParseQuality::Clean)
+                        {
+                            reconciliation_safe = false;
+                        }
                         let entity_count = fragment.functions.len()
                             + fragment.classes.len()
                             + fragment.imports.len()
@@ -1106,6 +1145,8 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
     // building them).
     {
         let mut projection = (*graph.snapshot()).clone();
+        let mut current_concept_ids = HashSet::new();
+        let mut concept_flush_succeeded = true;
         // Dotted-name resolution (resolve_imports, resolve_all_calls) is O(1)
         // per lookup once this suffix index is built; without it the cascade
         // scans every module per import (v0.8 P1: 8.3s of the 46s benchmark).
@@ -1119,7 +1160,11 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
         // Concept JSON v2 flush — post-cascade, pre-edge (FK order).
         if let Some(ref store) = graph.store {
             let v2 = crate::storage::build_v2_concepts_all(&projection);
-            let _ = store.upsert_concepts_bulk(&v2);
+            current_concept_ids = v2.iter().map(|concept| concept.id.clone()).collect();
+            if let Err(e) = store.upsert_concepts_bulk(&v2) {
+                concept_flush_succeeded = false;
+                eprintln!("[diag] v2 concept flush failed: {e:?}");
+            }
         }
         // Persist resolved edges to Macrame store
         if let Err(e) = graph.persist_edges(&projection) {
@@ -1164,6 +1209,19 @@ fn analyze_inner(root: &str, create_store: bool, extra_excludes: &[String]) -> A
                     Ok(_) => {}
                     Err(e) => eprintln!("[diag] non-canonical retraction failed: {e:?}"),
                 }
+                // Retire stale canonical rows (notably line-addressed imports)
+                // only when every source was walked/read/extracted and the v2
+                // projection flush succeeded. Otherwise missing data in this
+                // run is not evidence that a ledger concept was deleted.
+                if reconciliation_safe && failures.is_empty() && concept_flush_succeeded {
+                    match store.retire_stale_file_concepts(&current_concept_ids) {
+                        Ok(n) if n > 0 => eprintln!(
+                            "[coderadar] retired {n} stale file concept(s) absent from the fresh index"
+                        ),
+                        Ok(_) => {}
+                        Err(e) => eprintln!("[diag] stale concept retraction failed: {e:?}"),
+                    }
+                }
                 // R2-2: edge retraction. The persist above is assert-only
                 // and the concept retractions only close edges of removed
                 // entities, so an edge whose endpoints both survive but
@@ -1337,7 +1395,31 @@ fn remove_file(file_path: &str) -> PyResult<PyObject> {
         // F14: remove by the canonical form or the walk-form entry is
         // never found (forward-slash input vs `\` -joined stored ids).
         let canonical = crate::graph::module_resolution::canonical_file_form(file_path);
-        let removed = graph.remove_file(&canonical);
+        let mut removed = graph.remove_file(&canonical);
+        // The in-memory projection may have lost track of concepts written by
+        // an interrupted or older indexer. Use the indexed `extra.file_path`
+        // metadata to retire any such current rows as part of file deletion.
+        if let Some(store) = graph.store.as_ref() {
+            match store.live_concept_ids_for_file(&canonical) {
+                Ok(ledger_ids) => {
+                    let known: std::collections::HashSet<&str> =
+                        removed.iter().map(String::as_str).collect();
+                    let stale: Vec<String> = ledger_ids
+                        .into_iter()
+                        .filter(|id| !known.contains(id.as_str()))
+                        .collect();
+                    if !stale.is_empty() {
+                        match store.retire_entities(&stale) {
+                            Ok(_) => removed.extend(stale),
+                            Err(e) => eprintln!(
+                                "[diag] failed to retire stale concepts for {canonical}: {e:?}"
+                            ),
+                        }
+                    }
+                }
+                Err(e) => eprintln!("[diag] indexed concept lookup failed for {canonical}: {e:?}"),
+            }
+        }
         let dict = PyDict::new(py);
         dict.set_item("entities_removed", removed.len())?;
         dict.set_item("removed_ids", removed)?;