Added Rust benchmark suite

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/metadata.json

@@ -0,0 +1,18 @@
+{
+  "seed_type": "bug-seeded",
+  "expected_files": ["src/lib.rs"],
+  "reference_patch_size_lines": 2,
+  "primary_skill": "rust-debugging",
+  "secondary_skills": ["cargo-test", "chrono"],
+  "name": "rust_bugfix_001_chrono_off_by_one",
+  "difficulty": "easy",
+  "runtime_stressors": ["fault_localization", "test_interpretation"],
+  "task_type": "single_file_bugfix",
+  "requires_repo_navigation": false,
+  "requires_multi_step_reasoning": false,
+  "has_false_fix_trap": false,
+  "requires_retry_recovery": false,
+  "likely_tool_sequence": ["read_files", "edit_files", "run_tests"],
+  "evaluation_focus": ["correctness", "deterministic_tests"],
+  "benchmark_bucket": "baseline"
+}

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/prompt.txt

@@ -0,0 +1 @@
+The CLI command `report --date today` returns tomorrow's date in some timezones. Fix the bug and verify the behavior.

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/repo/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "date-resolver"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+chrono = "0.4"
+
+[[test]]
+name = "test_fix"
+path = "tests/test_fix.rs"

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/repo/src/lib.rs

@@ -0,0 +1,6 @@
+use chrono::{DateTime, Duration, NaiveDate, Utc};
+
+pub fn resolve_today(now: DateTime<Utc>) -> NaiveDate {
+    let shifted = now + Duration::hours(24);
+    shifted.date_naive()
+}

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/repo/tests/test_fix.rs

@@ -0,0 +1,24 @@
+use chrono::{TimeZone, Utc};
+use date_resolver::resolve_today;
+
+#[test]
+fn test_resolve_today_returns_correct_date() {
+    let now = Utc.with_ymd_and_hms(2024, 3, 15, 10, 0, 0).unwrap();
+    let result = resolve_today(now);
+    assert_eq!(
+        result.to_string(),
+        "2024-03-15",
+        "resolve_today should return today's date, not tomorrow's"
+    );
+}
+
+#[test]
+fn test_resolve_today_near_midnight() {
+    let now = Utc.with_ymd_and_hms(2024, 6, 1, 23, 30, 0).unwrap();
+    let result = resolve_today(now);
+    assert_eq!(
+        result.to_string(),
+        "2024-06-01",
+        "resolve_today near midnight should still return the current date"
+    );
+}

benchmark_tasks/villani_rust_v1/rust_bugfix_001_chrono_off_by_one/task.yaml

@@ -0,0 +1,28 @@
+id: rust_bugfix_001_chrono_off_by_one
+family: bugfix
+difficulty: easy
+language: rust
+max_minutes: 5
+max_files_touched: 3
+expected_artifacts:
+- patch
+visible_verification:
+- cargo test --test test_fix -q
+hidden_verification:
+- cargo test -q
+success_policy:
+  require_visible_pass: true
+  require_hidden_pass: true
+  fail_on_timeout: true
+  fail_on_repo_dirty_outside_allowlist: true
+allowlist_paths:
+- target/
+- src/
+- tests/
+allowed_paths:
+- target/
+- src/
+- tests/
+expected_touched_max: 3
+task_type: single_file_bugfix
+benchmark_track: core

benchmark_tasks/villani_rust_v1/rust_bugfix_002_env_config_precedence/metadata.json

@@ -0,0 +1,18 @@
+{
+  "seed_type": "bug-seeded",
+  "expected_files": ["src/lib.rs"],
+  "reference_patch_size_lines": 3,
+  "primary_skill": "rust-debugging",
+  "secondary_skills": ["cargo-test", "std-env"],
+  "name": "rust_bugfix_002_env_config_precedence",
+  "difficulty": "easy",
+  "runtime_stressors": ["fault_localization", "test_interpretation"],
+  "task_type": "single_file_bugfix",
+  "requires_repo_navigation": false,
+  "requires_multi_step_reasoning": false,
+  "has_false_fix_trap": false,
+  "requires_retry_recovery": false,
+  "likely_tool_sequence": ["read_files", "edit_files", "run_tests"],
+  "evaluation_focus": ["correctness", "deterministic_tests"],
+  "benchmark_bucket": "baseline"
+}

benchmark_tasks/villani_rust_v1/rust_bugfix_002_env_config_precedence/prompt.txt

@@ -0,0 +1 @@
+When both an environment variable and a config file value are present, the application should prefer the environment variable. Currently the config file value always wins regardless. Fix the precedence bug and verify it.