fix(eval): resolve assertion field access regression in merged objects (#44)

Move run_assertions() before the Pending cache marker in get_idx so
assertion field accesses don't see false Pending markers for fields that
aren't actually being computed yet.

The regression was introduced by the upstream jrsonnet self/super refactor
(faca88a) combined with the error propagation change (ed68a22). When an
assertion in a super core of a merged object accessed a field via self
(e.g. `{ assert self.used > 0 } + { used: 1 }`), the Pending marker set
before run_assertions() caused the field lookup to incorrectly assume
circular recursion and fall back to searching only super cores, which
didn't have the field.

Also fix suggest_object_fields to skip exact matches instead of panicking
when string pooling produces duplicate interned strings under concurrent
evaluation.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: julienduchesne

cmds/rtk/testdata/test-export-used-field-regression/main.jsonnet

@@ -0,0 +1,14 @@
+local cfg = ({ assert self.used > 0 : 'used must be positive' } + { used: 1 });
+
+{
+  config: {
+    apiVersion: 'v1',
+    kind: 'ConfigMap',
+    metadata: {
+      name: 'used-field-regression',
+    },
+    data: {
+      used: std.toString(cfg.used),
+    },
+  },
+}

cmds/rtk/testdata/test-export-used-field-regression/spec.json

@@ -0,0 +1,10 @@
+{
+  "apiVersion": "tanka.dev/v1alpha1",
+  "kind": "Environment",
+  "metadata": {
+    "name": "test-export-used-field-regression"
+  },
+  "spec": {
+    "namespace": "default"
+  }
+}

cmds/rtk/tests/export_test.rs

@@ -2174,3 +2174,39 @@ fn test_export_tla_function_with_overrides() {
 		deployment_content
 	);
 }
+
+/// Minimal anonymized regression case for "no such field: used".
+///
+/// This reproduces a known evaluator regression where an assertion that reads
+/// `self.used` from a merged object can fail with:
+/// `no such field: used`.
+#[test]
+fn regression_minimal_used_field_export() {
+	let temp_dir = tempfile::TempDir::new().unwrap();
+	let output_dir = temp_dir.path();
+
+	let opts = ExportOpts {
+		output_dir: output_dir.to_path_buf(),
+		extension: "yaml".to_string(),
+		parallelism: 1,
+		recursive: false,
+		eval_opts: EvalOpts::default(),
+		..Default::default()
+	};
+
+	let result = export(
+		&[testdata_path("test-export-used-field-regression")
+			.to_string_lossy()
+			.to_string()],
+		opts,
+	)
+	.expect("export should succeed");
+
+	assert_eq!(result.successful, 1);
+	assert_eq!(result.failed, 0);
+
+	check_files(
+		output_dir,
+		&["manifest.json", "v1.ConfigMap-used-field-regression.yaml"],
+	);
+}

crates/jrsonnet-evaluator/src/error.rs

@@ -97,7 +97,11 @@ pub(crate) fn suggest_object_fields(v: &ObjValue, key: IStr) -> Vec<IStr> {
 		if conf < 0.8 {
 			continue;
 		}
-		assert!(field.as_str() != key.as_str(), "looks like string pooling failure, please write any info regarding this crash to https://github.com/CertainLach/jrsonnet/issues/113, thanks!");
+		// Skip exact match: don't suggest the key itself. Compare by string content so we
+		// don't panic when string pooling fails (e.g. under heavy load with many envs).
+		if field.as_str() == key.as_str() {
+			continue;
+		}
 
 		heap.push((conf, field));
 	}