poimen-memory/tests/it_labeler.rs

use mem_llm::{EvidenceLabel, make_label_prompt, parse_label_response, fits_context_budget};

/// M5.1 Integration Tests — Evidence Labeler
/// 
/// Verifies the evidence labeling pipeline:
/// - Labels keyed by chunk_sha (survives re-chunking)
/// - Context budget checked (16384 token limit)
/// - Justifications preserved for calibration
/// - Prompts have no tools (reasoning model requirement)

#[test]
fn a1_one_label_per_chunk() {
    let chunks = vec![
        "chunk1", "chunk2", "chunk3", "chunk4", "chunk5"
    ];
    
    let labels: Vec<EvidenceLabel> = chunks
        .iter()
        .enumerate()
        .map(|(i, sha)| {
            EvidenceLabel::new(
                sha.to_string(),
                i,
                i % 2 == 0,  // Alternate yes/no
                "Test justification".to_string(),
            )
        })
        .collect();
    
    assert_eq!(labels.len(), 5, "One label per chunk");
    assert!(labels.iter().all(|l| !l.chunk_sha.is_empty()), "All have sha");
    
    // No duplicates
    let mut shas = labels.iter().map(|l| &l.chunk_sha).collect::<Vec<_>>();
    let original_len = shas.len();
    shas.sort();
    shas.dedup();
    assert_eq!(shas.len(), original_len, "No duplicate labels");
}

#[test]
fn a2_keyed_by_sha() {
    // Labels are keyed on chunk_sha, not turn number
    let label1 = EvidenceLabel::new(
        "abc123".to_string(),
        7,
        true,
        "Contains evidence".to_string(),
    );
    
    let label2 = EvidenceLabel::new(
        "abc123".to_string(),
        5,  // Different turn number
        true,
        "Contains evidence".to_string(),
    );
    
    // Same chunk_sha = same label, regardless of turn order
    assert_eq!(label1.chunk_sha, label2.chunk_sha);
    // (In practice, we'd deduplicate by sha)
}

#[test]
fn a3_context_budget_respected() {
    let question = "What causes the timeout?";
    let chunk = "The database query lacks an index, causing sequential scans that take 30 seconds.";
    
    let prompt = make_label_prompt(question, chunk);
    
    // Should fit in reasoning model's 16384 limit
    assert!(fits_context_budget(&prompt, 64, 16384), "Reasonable chunk should fit");
}

#[test]
fn a4_justification_kept() {
    let responses = vec![
        "yes\nThis chunk directly answers the question about timeouts.",
        "no\nThis chunk discusses unrelated infrastructure.",
    ];
    
    for response in responses {
        let (label, why) = parse_label_response(response)
            .expect("Should parse label response");
        
        assert!(!why.is_empty(), "Justification should be preserved");
        assert!(why.len() > 10, "Justification should be a full sentence");
    }
}

#[test]
fn a5_label_structure() {
    let label = EvidenceLabel::new(
        "sha256abc".to_string(),
        12,
        true,
        "This chunk contains the evidence".to_string(),
    );
    
    assert_eq!(label.chunk_sha, "sha256abc");
    assert_eq!(label.t, 12);
    assert!(label.label);
    assert_eq!(label.why, "This chunk contains the evidence");
    assert_eq!(label.model, "reasoning");
    assert!(!label.ts.is_empty());
}

#[test]
fn a6_prompt_no_tools_field() {
    let prompt = make_label_prompt(
        "What is the issue?",
        "The service is down.",
    );
    
    // Reasoning model rejects tools - prompt should never contain them
    assert!(
        !prompt.contains("tools"),
        "Labeling prompt must not include tools field"
    );
    assert!(
        !prompt.contains("function_calls"),
        "Labeling prompt must not include function calls"
    );
}

#[test]
fn a7_parsing_handles_variations() {
    let variations = vec![
        ("YES\nThis is evidence", true),
        ("no\nThis is not evidence", false),
        ("Yes, definitely\nEvidence present", true),
        ("No, unrelated", false),
    ];
    
    for (response, expected_label) in variations {
        let (label, why) = parse_label_response(response)
            .expect("Should parse");
        
        assert_eq!(label, expected_label);
        assert!(!why.is_empty());
    }
}

#[test]
fn a8_empty_prompt_safe() {
    let prompt = make_label_prompt("", "");
    
    // Should still be valid (just asking labeler to work with nothing)
    assert!(!prompt.is_empty());
}

#[test]
fn a9_large_chunk_exceeds_budget() {
    let question = "What happened?";
    let huge_chunk = "x".repeat(100000);
    
    let prompt = make_label_prompt(question, &huge_chunk);
    
    // Should NOT fit in context
    assert!(!fits_context_budget(&prompt, 64, 16384));
}

#[test]
fn a10_label_rate_summary() {
    let labels = vec![
        EvidenceLabel::new("a".to_string(), 1, true, "yes".to_string()),
        EvidenceLabel::new("b".to_string(), 2, false, "no".to_string()),
        EvidenceLabel::new("c".to_string(), 3, true, "yes".to_string()),
        EvidenceLabel::new("d".to_string(), 4, false, "no".to_string()),
        EvidenceLabel::new("e".to_string(), 5, true, "yes".to_string()),
    ];
    
    let positive = labels.iter().filter(|l| l.label).count();
    let rate = positive as f32 / labels.len() as f32;
    
    assert_eq!(positive, 3, "3 out of 5 labeled as evidence");
    assert!((rate - 0.6).abs() < 0.01, "Label rate should be 60%");
}

#[test]
fn a11_evidence_label_serde() {
    let label = EvidenceLabel::new(
        "abc123def456".to_string(),
        7,
        true,
        "Contains direct evidence of the bug".to_string(),
    );
    
    // Should be serializable to JSON (for JSONL output)
    let json = serde_json::to_string(&label)
        .expect("Should serialize");
    
    assert!(json.contains("abc123def456"));
    assert!(json.contains("true"));
    assert!(json.contains("evidence"));
    
    // Should deserialize back
    let deserialized: EvidenceLabel = serde_json::from_str(&json)
        .expect("Should deserialize");
    
    assert_eq!(deserialized.chunk_sha, label.chunk_sha);
    assert_eq!(deserialized.label, label.label);
}