ai_memory/

bench.rs

// Copyright 2026 AlphaOne LLC
// SPDX-License-Identifier: Apache-2.0

//! Pillar 3 / Stream E — `ai-memory bench` workload runner.
//!
//! Measures hot-path operations against the budgets published in
//! `PERFORMANCE.md` and returns p50/p95/p99 latencies plus a pass/fail
//! verdict per operation. The CI guard (Stream F) enforces the same
//! 10% p95 tolerance documented in `PERFORMANCE.md`.
//!
//! Coverage in this build:
//! - Embedding-free CRUD: `memory_store` (no embedding), `memory_search`
//!   (FTS5), `memory_recall` (hot, depth=1).
//! - Knowledge-graph traversal:
//!     - `memory_kg_query` (depth=1) and `memory_kg_timeline` against a
//!       fan-out fixture (50 sources × 4 outbound links each, every
//!       link `valid_from`-stamped).
//!     - `memory_kg_query` (depth=3, depth=5) against a chain fixture
//!       (50 chains × 5 hops each = 300 memories + 250 links). depth=3
//!       hits the "depth ≤ 3" 100 ms budget bucket; depth=5 hits the
//!       "depth ≤ 5" 250 ms tail-case bucket.
//!
//! Both fixtures live in the same in-process disposable `SQLite` — no
//! external service required.
//!
//! Embedding-bound paths (`memory_store` with embedding,
//! `memory_recall` cold/full hybrid) still require an embedder process
//! and are tracked as follow-up Stream E work — they don't belong on
//! the hot path of a `cargo test` invocation.

use crate::models::ConfidenceSource;
use anyhow::{Context, Result};
use rusqlite::Connection;
use serde::{Deserialize, Serialize};
use std::path::Path;
use std::time::{Duration, Instant};

use crate::db;
use crate::models::{Memory, Tier};

/// CI guard tolerance — measured p95 may exceed budget by this factor
/// before the run is marked `Fail`. Mirrors `PERFORMANCE.md`.
pub const P95_TOLERANCE: f64 = 1.10;

/// macOS-runner budget multiplier (issue #1193).
///
/// Apple's `macos-latest` GHA runner pool has substantially higher
/// I/O scheduling variance and cold-start latency than `ubuntu-latest`.
/// `tests/integration.rs::test_cli_bench_emits_json_with_seven_results_and_passes_budget`
/// drives `ai-memory bench --iterations 5` end-to-end and asserts the
/// process exits 0 — at the small iteration count the macOS tail can
/// blow the absolute `target_p95_ms` budgets even when the underlying
/// code is healthy. Per #1193 "Proposed fix" option 1 (preferred):
/// apply a centralized multiplier inside the runner-effective budget
/// path so the pass/fail verdict is platform-aware while the canonical
/// `target_p95_ms` reported in the JSON envelope still reflects the
/// PERFORMANCE.md numbers (unchanged for dashboards / regression
/// trackers). Multiplier of 3 mirrors the same headroom applied to the
/// timing-sensitive hooks tests under the same issue.
#[cfg(target_os = "macos")]
pub const MACOS_BUDGET_MULT: f64 = 3.0;
#[cfg(not(target_os = "macos"))]
pub const MACOS_BUDGET_MULT: f64 = 1.0;

/// Default seeded namespace for the bench workload.
pub const BENCH_NAMESPACE: &str = "ai-memory-bench";

/// Default workload size — keep small enough for `cargo test`, large
/// enough that p99 has signal.
pub const DEFAULT_ITERATIONS: usize = 200;

/// Default warmup iterations discarded from the percentile sample.
pub const DEFAULT_WARMUP: usize = 20;

/// Hard ceiling on `--iterations` — bounds bench wall-clock on a
/// mistyped flag.
pub const MAX_ITERATIONS: usize = 100_000;

/// Hard ceiling on `--warmup` iterations.
pub const MAX_WARMUP: usize = 10_000;

/// Hard ceiling on `--regression-threshold` (percent) — values above
/// this are clamped; a 1000% allowance already means "no gate".
pub const MAX_REGRESSION_THRESHOLD_PCT: f64 = 1000.0;

/// #1579 B8 — canonical corpus scale (rows) for the scale-gate run
/// (`ai-memory bench --scale 10000`). The P1 perf-audit proved the
/// default workload (~500 rows after per-op seeding) cannot see
/// corpus-scale budget blowouts (recall p95 361 ms vs the 50 ms budget
/// at 100k rows was invisible to the built-in bench).
pub const CI_SCALE_GATE_ROWS: usize = 10_000;

/// #1579 B8 — hard ceiling on `--scale` rows. Bounds seeding wall-clock
/// + RAM on a mistyped flag (1M rows ≈ the largest corpus the perf
/// audit exercised).
pub const MAX_SCALE: usize = 1_000_000;

/// #1579 B8 — one row of the per-scale p95 budget table published in
/// `PERFORMANCE.md` §"Corpus-scale budgets". Only the three
/// corpus-sensitive operations carry scale-specific budgets; the KG
/// operations run against fixed-size fixtures (50×4 fan-out, 50×5
/// chains) whose cost is independent of the seeded corpus scale, so
/// they keep their canonical budgets at every scale.
#[derive(Debug, Clone, Copy)]
pub struct ScaleBudgets {
    /// Seeded corpus rows this row's budgets apply to (upper bound —
    /// a requested scale selects the first table row whose `scale` is
    /// `>=` the request).
    pub scale: usize,
    /// `memory_store` (no embedding) p95 budget, ms.
    pub store_no_embedding_ms: f64,
    /// `memory_search` (FTS5) p95 budget, ms.
    pub search_fts_ms: f64,
    /// `memory_recall` (hot, keyword) p95 budget, ms.
    pub recall_hot_ms: f64,
}

/// #1579 B8 — the per-scale budget table (SSOT; `PERFORMANCE.md`
/// §"Corpus-scale budgets" narrates these numbers and the
/// `operation_scale_targets_match_performance_md` test pins them).
///
/// 10k-row budgets were pinned from a measured release-build run on
/// this branch (`ai-memory bench --scale 10000`, Linux x86_64) with
/// ≥50% headroom over the measurement, capped at the operator-approved
/// conservative ceilings from the #1579 remediation plan (store ≤120,
/// recall ≤80, search ≤60).
pub const SCALE_BUDGETS: &[ScaleBudgets] = &[ScaleBudgets {
    scale: CI_SCALE_GATE_ROWS,
    store_no_embedding_ms: 120.0,
    search_fts_ms: 60.0,
    recall_hot_ms: 80.0,
}];

/// #1579 B8 — resolve the budget row for a requested scale: the first
/// table row whose `scale >= requested`, else the largest pinned row
/// (best-effort; pin a new table row before gating larger scales).
#[must_use]
pub fn scale_budgets_for(requested: usize) -> ScaleBudgets {
    for row in SCALE_BUDGETS {
        if row.scale >= requested {
            return *row;
        }
    }
    *SCALE_BUDGETS
        .last()
        .expect("SCALE_BUDGETS table must be non-empty")
}

/// Default tolerance applied when comparing a fresh run against a
/// `--baseline` JSON file: a measured p95 may grow by this percentage
/// before the run is flagged as a regression. Independent of
/// [`P95_TOLERANCE`] (which guards against the absolute budget). The
/// baseline guard catches drift that stays inside the absolute budget
/// but trends in the wrong direction across releases.
pub const DEFAULT_REGRESSION_THRESHOLD_PCT: f64 = 10.0;

/// Hot-path operations covered by this iteration of the bench tool.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum Operation {
    /// `memory_store` without embedding — pure `SQLite` write path.
    StoreNoEmbedding,
    /// `memory_search` — FTS5 keyword baseline.
    SearchFts,
    /// `memory_recall` hot path, depth=1 (no hierarchy expansion).
    RecallHot,
    /// `memory_kg_query` recursive-CTE traversal at depth=1 (the
    /// shallowest path through the depth ≤ 3 budget bucket).
    KgQueryDepth1,
    /// `memory_kg_query` recursive-CTE traversal at depth=3 (the
    /// deepest path inside the "depth ≤ 3" 100 ms budget bucket). Driven
    /// against a chain fixture so the recursive CTE actually visits
    /// three hops per query.
    KgQueryDepth3,
    /// `memory_kg_query` recursive-CTE traversal at depth=5 (the tail
    /// case for the "depth ≤ 5" 250 ms budget bucket). Driven against
    /// the same chain fixture as depth=3.
    KgQueryDepth5,
    /// `memory_kg_timeline` — ordered timeline for a single source.
    KgTimeline,
}

impl Operation {
    #[must_use]
    pub fn label(self) -> &'static str {
        match self {
            Self::StoreNoEmbedding => "memory_store (no embedding)",
            Self::SearchFts => "memory_search (FTS5)",
            Self::RecallHot => "memory_recall (hot, depth=1)",
            Self::KgQueryDepth1 => "memory_kg_query (depth=1)",
            Self::KgQueryDepth3 => "memory_kg_query (depth=3)",
            Self::KgQueryDepth5 => "memory_kg_query (depth=5)",
            Self::KgTimeline => crate::mcp::registry::tool_names::MEMORY_KG_TIMELINE,
        }
    }

    /// p95 budget in milliseconds, sourced from `PERFORMANCE.md`.
    ///
    /// `KgQueryDepth1` and `KgQueryDepth3` both fall in the
    /// "depth ≤ 3" (100 ms) bucket; `KgQueryDepth5` is the tail case
    /// at "depth ≤ 5" (250 ms). `SearchFts` and `KgTimeline` happen to
    /// share the same numeric budget as the depth ≤ 3 bucket despite
    /// belonging to different table rows in `PERFORMANCE.md`.
    ///
    /// This is the canonical published budget; the runner-effective
    /// pass/fail verdict uses [`effective_target_p95_ms`] which
    /// applies the [`MACOS_BUDGET_MULT`] platform multiplier on top.
    #[must_use]
    #[allow(clippy::match_same_arms)]
    pub fn target_p95_ms(self) -> f64 {
        match self {
            Self::StoreNoEmbedding => 20.0,
            Self::SearchFts => 100.0,
            Self::RecallHot => 50.0,
            Self::KgQueryDepth1 => 100.0,
            Self::KgQueryDepth3 => 100.0,
            Self::KgQueryDepth5 => 250.0,
            Self::KgTimeline => 100.0,
        }
    }

    /// Runner-effective p95 budget — equal to [`target_p95_ms`] on
    /// Linux/Windows, but multiplied by [`MACOS_BUDGET_MULT`] on
    /// macOS targets per issue #1193. The pass/fail verdict in the
    /// CLI bench tool uses this value; the JSON envelope's
    /// `target_p95_ms` field continues to report the canonical
    /// PERFORMANCE.md number so regression dashboards stay stable.
    #[must_use]
    pub fn effective_target_p95_ms(self) -> f64 {
        self.target_p95_ms() * MACOS_BUDGET_MULT
    }

    /// #1579 B8 — canonical p95 budget at a given corpus scale.
    /// `None` (the default workload) keeps the legacy
    /// [`Self::target_p95_ms`] budgets byte-for-byte. `Some(rows)`
    /// swaps in the [`SCALE_BUDGETS`] row for the three
    /// corpus-sensitive operations; the KG operations keep their
    /// canonical budgets because their fixtures are scale-independent
    /// (see [`ScaleBudgets`]).
    #[must_use]
    pub fn target_p95_ms_at_scale(self, scale: Option<usize>) -> f64 {
        let Some(rows) = scale else {
            return self.target_p95_ms();
        };
        let budgets = scale_budgets_for(rows);
        match self {
            Self::StoreNoEmbedding => budgets.store_no_embedding_ms,
            Self::SearchFts => budgets.search_fts_ms,
            Self::RecallHot => budgets.recall_hot_ms,
            Self::KgQueryDepth1 | Self::KgQueryDepth3 | Self::KgQueryDepth5 | Self::KgTimeline => {
                self.target_p95_ms()
            }
        }
    }

    /// #1579 B8 — runner-effective sibling of
    /// [`Self::target_p95_ms_at_scale`] (applies the #1193 macOS
    /// multiplier, same as [`Self::effective_target_p95_ms`]).
    #[must_use]
    pub fn effective_target_p95_ms_at_scale(self, scale: Option<usize>) -> f64 {
        self.target_p95_ms_at_scale(scale) * MACOS_BUDGET_MULT
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum Status {
    Pass,
    Fail,
}

#[derive(Debug, Clone, Serialize)]
pub struct OperationResult {
    pub operation: Operation,
    /// Pretty label, duplicated for JSON consumers.
    pub label: &'static str,
    pub target_p95_ms: f64,
    pub measured_p50_ms: f64,
    pub measured_p95_ms: f64,
    pub measured_p99_ms: f64,
    pub samples: usize,
    pub status: Status,
}

#[derive(Debug, Clone)]
pub struct BenchConfig {
    pub iterations: usize,
    pub warmup: usize,
    pub namespace: String,
    /// #1579 B8 — corpus scale. `None` keeps the legacy default
    /// workload (~500 rows after per-op seeding) and the legacy
    /// budgets; `Some(rows)` seeds a scratch corpus of `rows` rows
    /// into the bench namespace before the operations run and gates
    /// the verdict against the [`SCALE_BUDGETS`] table instead.
    pub scale: Option<usize>,
}

impl Default for BenchConfig {
    fn default() -> Self {
        Self {
            iterations: DEFAULT_ITERATIONS,
            warmup: DEFAULT_WARMUP,
            namespace: BENCH_NAMESPACE.to_string(),
            scale: None,
        }
    }
}

/// Run the bench workload and return per-operation results.
///
/// Each operation seeds its own data inside the supplied connection so
/// callers can hand in either a fresh in-memory DB (for tests) or a
/// disposable on-disk DB (for the CLI).
///
/// # Errors
///
/// Returns the underlying [`db`] error if any of the seeded inserts
/// or queries fail.
pub fn run(conn: &Connection, config: &BenchConfig) -> Result<Vec<OperationResult>> {
    // #1579 B8 — seed the scratch corpus FIRST so every operation
    // below (FTS5 search, hybrid-keyword recall, the store upsert
    // probe) runs against a table of ~`scale` rows, not the ~500-row
    // default that hid the 100k-corpus budget blowouts from the P1
    // audit. The corpus shares the bench namespace and the
    // `topic-N / category-M` vocabulary the search/recall queries use,
    // so the queries genuinely scan it.
    if let Some(rows) = config.scale {
        seed_corpus(conn, &config.namespace, "scale", rows)?;
    }
    let store = run_store_no_embedding(conn, config)?;
    let search = run_search_fts(conn, config)?;
    let recall = run_recall_hot(conn, config)?;
    let kg_sources = seed_kg_fixture(conn, &config.namespace)?;
    let kg_query = run_kg_query_depth1(conn, config, &kg_sources)?;
    let kg_chain_sources = seed_kg_chain_fixture(conn, &config.namespace)?;
    let kg_query_d3 =
        run_kg_query_chain(conn, config, &kg_chain_sources, Operation::KgQueryDepth3, 3)?;
    let kg_query_d5 =
        run_kg_query_chain(conn, config, &kg_chain_sources, Operation::KgQueryDepth5, 5)?;
    let kg_timeline = run_kg_timeline(conn, config, &kg_sources)?;
    Ok(vec![
        store,
        search,
        recall,
        kg_query,
        kg_query_d3,
        kg_query_d5,
        kg_timeline,
    ])
}

fn run_store_no_embedding(conn: &Connection, config: &BenchConfig) -> Result<OperationResult> {
    let total = config.warmup + config.iterations;
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..total {
        let mem = synth_memory(&config.namespace, i, "store");
        let start = Instant::now();
        db::insert(conn, &mem)?;
        let elapsed = start.elapsed();
        if i >= config.warmup {
            samples.push(elapsed);
        }
    }
    Ok(percentile_summary(
        Operation::StoreNoEmbedding,
        &samples,
        config.scale,
    ))
}

fn run_search_fts(conn: &Connection, config: &BenchConfig) -> Result<OperationResult> {
    seed_corpus(conn, &config.namespace, "search", 200)?;
    let total = config.warmup + config.iterations;
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..total {
        let query = format!("topic-{}", i % 50);
        let start = Instant::now();
        let _ = db::search(
            conn,
            &query,
            Some(&config.namespace),
            None,
            10,
            None,
            None,
            None,
            None,
            None,
            None,
            false,
        )?;
        let elapsed = start.elapsed();
        if i >= config.warmup {
            samples.push(elapsed);
        }
    }
    Ok(percentile_summary(
        Operation::SearchFts,
        &samples,
        config.scale,
    ))
}

fn run_recall_hot(conn: &Connection, config: &BenchConfig) -> Result<OperationResult> {
    seed_corpus(conn, &config.namespace, "recall", 200)?;
    let warmup_query = "topic 0 category 0";
    for _ in 0..config.warmup {
        let _ = db::recall(
            conn,
            warmup_query,
            Some(&config.namespace),
            10,
            None,
            None,
            None,
            0,
            0,
            None,
            None,
            false,
            None,
        )?;
    }
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..config.iterations {
        let query = format!("topic {} category {}", i % 50, i % 10);
        let start = Instant::now();
        let _ = db::recall(
            conn,
            &query,
            Some(&config.namespace),
            10,
            None,
            None,
            None,
            0,
            0,
            None,
            None,
            false,
            None,
        )?;
        samples.push(start.elapsed());
    }
    Ok(percentile_summary(
        Operation::RecallHot,
        &samples,
        config.scale,
    ))
}

/// Source memory IDs returned from [`seed_kg_fixture`]. Each source has
/// `KG_FIXTURE_LINKS_PER_SOURCE` outbound links — the bench drives both
/// `kg_query` and `kg_timeline` against the same fixture.
const KG_FIXTURE_SOURCES: usize = 50;
const KG_FIXTURE_LINKS_PER_SOURCE: usize = 4;

/// Linear-chain fixture geometry for the depth=3 / depth=5 runners.
/// `KG_CHAIN_FIXTURE_CHAINS` chains × `KG_CHAIN_FIXTURE_HOPS` hops yields
/// `chains * (hops + 1)` memories and `chains * hops` links — so 50 × 5
/// matches the fan-out fixture's order of magnitude (300 memories +
/// 250 links). depth=5 reaches every node in a chain; depth=3 reaches
/// the first three follow-on hops.
const KG_CHAIN_FIXTURE_CHAINS: usize = 50;
const KG_CHAIN_FIXTURE_HOPS: usize = 5;

fn run_kg_query_depth1(
    conn: &Connection,
    config: &BenchConfig,
    sources: &[String],
) -> Result<OperationResult> {
    debug_assert!(
        !sources.is_empty(),
        "kg_query bench requires a seeded fixture"
    );
    let total = config.warmup + config.iterations;
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..total {
        let src = &sources[i % sources.len()];
        let start = Instant::now();
        let _ = db::kg_query(conn, src, 1, None, None, None, false)?;
        let elapsed = start.elapsed();
        if i >= config.warmup {
            samples.push(elapsed);
        }
    }
    Ok(percentile_summary(
        Operation::KgQueryDepth1,
        &samples,
        config.scale,
    ))
}

fn run_kg_query_chain(
    conn: &Connection,
    config: &BenchConfig,
    sources: &[String],
    operation: Operation,
    max_depth: usize,
) -> Result<OperationResult> {
    debug_assert!(
        !sources.is_empty(),
        "kg_query chain bench requires a seeded fixture"
    );
    let total = config.warmup + config.iterations;
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..total {
        let src = &sources[i % sources.len()];
        let start = Instant::now();
        let _ = db::kg_query(conn, src, max_depth, None, None, None, false)?;
        let elapsed = start.elapsed();
        if i >= config.warmup {
            samples.push(elapsed);
        }
    }
    Ok(percentile_summary(operation, &samples, config.scale))
}

fn run_kg_timeline(
    conn: &Connection,
    config: &BenchConfig,
    sources: &[String],
) -> Result<OperationResult> {
    debug_assert!(
        !sources.is_empty(),
        "kg_timeline bench requires a seeded fixture"
    );
    let total = config.warmup + config.iterations;
    let mut samples = Vec::with_capacity(config.iterations);
    for i in 0..total {
        let src = &sources[i % sources.len()];
        let start = Instant::now();
        let _ = db::kg_timeline(conn, src, None, None, None)?;
        let elapsed = start.elapsed();
        if i >= config.warmup {
            samples.push(elapsed);
        }
    }
    Ok(percentile_summary(
        Operation::KgTimeline,
        &samples,
        config.scale,
    ))
}

/// Seed the in-process KG fixture: `KG_FIXTURE_SOURCES` source memories,
/// each with `KG_FIXTURE_LINKS_PER_SOURCE` outbound links to distinct
/// targets. Every link sets `valid_from` so `kg_timeline` (which skips
/// rows with NULL `valid_from`) sees the full corpus. Returns the source
/// IDs so the runners can hand them to `kg_query` / `kg_timeline`.
fn seed_kg_fixture(conn: &Connection, namespace: &str) -> Result<Vec<String>> {
    let mut sources = Vec::with_capacity(KG_FIXTURE_SOURCES);
    for s in 0..KG_FIXTURE_SOURCES {
        let src = synth_memory(namespace, s, "kg-src");
        // `db::insert` upserts on `(title, namespace)` and returns the
        // canonical id, which differs from `src.id` if the row already
        // exists. Use the returned id so the fixture remains correct
        // even when `run()` is invoked twice against the same conn.
        let src_id = db::insert(conn, &src)?;
        for t in 0..KG_FIXTURE_LINKS_PER_SOURCE {
            let target_idx = s * KG_FIXTURE_LINKS_PER_SOURCE + t;
            let tgt = synth_memory(namespace, target_idx, "kg-tgt");
            let tgt_id = db::insert(conn, &tgt)?;
            // `db::create_link` stamps `created_at` and `valid_from` to
            // the current wall clock — sufficient for `kg_timeline`
            // (which skips rows with NULL `valid_from`).
            db::create_link(
                conn,
                &src_id,
                &tgt_id,
                crate::models::MemoryLinkRelation::RelatedTo.as_str(),
            )?;
        }
        sources.push(src_id);
    }
    Ok(sources)
}

/// Seed the linear-chain KG fixture used by the depth=3 / depth=5
/// runners: `KG_CHAIN_FIXTURE_CHAINS` chains, each
/// `KG_CHAIN_FIXTURE_HOPS` links long. Every node and link uses titles
/// disjoint from the fan-out fixture's `kg-src` / `kg-tgt` prefixes, so
/// both fixtures coexist in the same connection without colliding on
/// the `(title, namespace)` upsert. Returns the source IDs (one per
/// chain) so the runners can drive `kg_query` against them.
fn seed_kg_chain_fixture(conn: &Connection, namespace: &str) -> Result<Vec<String>> {
    let mut sources = Vec::with_capacity(KG_CHAIN_FIXTURE_CHAINS);
    for c in 0..KG_CHAIN_FIXTURE_CHAINS {
        let mut prev_id = {
            let head = synth_memory(namespace, c, "kg-chain-src");
            db::insert(conn, &head)?
        };
        let chain_head_id = prev_id.clone();
        for h in 0..KG_CHAIN_FIXTURE_HOPS {
            let node_idx = c * KG_CHAIN_FIXTURE_HOPS + h;
            let next = synth_memory(namespace, node_idx, "kg-chain-node");
            let next_id = db::insert(conn, &next)?;
            db::create_link(
                conn,
                &prev_id,
                &next_id,
                crate::models::MemoryLinkRelation::RelatedTo.as_str(),
            )?;
            prev_id = next_id;
        }
        sources.push(chain_head_id);
    }
    Ok(sources)
}

fn seed_corpus(conn: &Connection, namespace: &str, prefix: &str, count: usize) -> Result<()> {
    for i in 0..count {
        let mem = synth_memory(namespace, i, prefix);
        db::insert(conn, &mem)?;
    }
    Ok(())
}

fn synth_memory(namespace: &str, i: usize, prefix: &str) -> Memory {
    let now = chrono::Utc::now().to_rfc3339();
    Memory {
        id: uuid::Uuid::new_v4().to_string(),
        tier: Tier::Long,
        namespace: namespace.to_string(),
        title: format!("bench-{prefix}-{i}"),
        content: format!(
            "bench memory {i} content about topic {} category {} for {prefix} workload",
            i % 50,
            i % 10
        ),
        tags: vec![],
        priority: i32::try_from((i % 9) + 1).unwrap_or(5),
        confidence: 1.0,
        source: "bench".to_string(),
        access_count: 0,
        created_at: now.clone(),
        updated_at: now,
        last_accessed_at: None,
        expires_at: None,
        metadata: serde_json::json!({"agent_id": "bench"}),
        reflection_depth: 0,
        memory_kind: crate::models::MemoryKind::Observation,
        entity_id: None,
        persona_version: None,
        citations: Vec::new(),
        source_uri: None,
        source_span: None,
        confidence_source: ConfidenceSource::CallerProvided,
        confidence_signals: None,
        confidence_decayed_at: None,
        version: 1,
    }
}

fn percentile_summary(
    operation: Operation,
    samples: &[Duration],
    // #1579 B8 — corpus scale of this run; selects the budget bucket.
    scale: Option<usize>,
) -> OperationResult {
    debug_assert!(
        !samples.is_empty(),
        "bench operation produced no samples; iterations must be > 0"
    );
    let mut sorted: Vec<f64> = samples.iter().map(duration_ms).collect();
    sorted.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
    let p50 = percentile(&sorted, 0.50);
    let p95 = percentile(&sorted, 0.95);
    let p99 = percentile(&sorted, 0.99);
    // #1579 B8 — both the reported target and the verdict budget come
    // from the scale-aware resolver; `scale == None` keeps the legacy
    // budgets byte-for-byte.
    let target = operation.target_p95_ms_at_scale(scale);
    // Per issue #1193: the pass/fail verdict uses the runner-effective
    // budget so the macOS GHA pool's higher I/O variance doesn't blow
    // a clean run. The reported `target_p95_ms` keeps the canonical
    // PERFORMANCE.md value so dashboards / baselines stay stable.
    let effective_target = operation.effective_target_p95_ms_at_scale(scale);
    let status = if p95 <= effective_target * P95_TOLERANCE {
        Status::Pass
    } else {
        Status::Fail
    };
    OperationResult {
        operation,
        label: operation.label(),
        target_p95_ms: target,
        measured_p50_ms: p50,
        measured_p95_ms: p95,
        measured_p99_ms: p99,
        samples: sorted.len(),
        status,
    }
}

fn duration_ms(d: &Duration) -> f64 {
    let secs = d.as_secs_f64();
    secs * 1000.0
}

#[allow(
    clippy::cast_precision_loss,
    clippy::cast_sign_loss,
    clippy::cast_possible_truncation
)]
fn percentile(sorted: &[f64], q: f64) -> f64 {
    if sorted.is_empty() {
        return 0.0;
    }
    if sorted.len() == 1 {
        return sorted[0];
    }
    let rank = q * (sorted.len() as f64 - 1.0);
    let lo = rank.floor() as usize;
    let hi = rank.ceil() as usize;
    if lo == hi {
        return sorted[lo];
    }
    let frac = rank - lo as f64;
    sorted[lo] + (sorted[hi] - sorted[lo]) * frac
}

/// Render a results table to a string in the same shape used in the
/// `PERFORMANCE.md` "Operator Self-Verification" example.
#[must_use]
pub fn render_table(results: &[OperationResult]) -> String {
    let mut out = String::new();
    out.push_str(
        "Operation                       Target (p95)   Measured (p95)   p50      p99      Status\n",
    );
    out.push_str(
        "─────────────────────────────────────────────────────────────────────────────────────────\n",
    );
    for r in results {
        let status_str = match r.status {
            Status::Pass => "PASS",
            Status::Fail => "FAIL",
        };
        // target budgets are documented as small integer ms; rounding
        // to the nearest int ms is what the table in PERFORMANCE.md
        // shows. Saturating cast guards against pathological future
        // changes to a non-integer or huge value.
        #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
        let target_ms = r.target_p95_ms.round() as i64;
        let line = format!(
            "{:<30}  < {:>4} ms       {:>7.1} ms       {:>5.1}    {:>5.1}    {}\n",
            r.label, target_ms, r.measured_p95_ms, r.measured_p50_ms, r.measured_p99_ms, status_str
        );
        out.push_str(&line);
    }
    out
}

/// Subset of [`OperationResult`] retained when loading a previous run
/// for `--baseline` comparison. Only the fields the regression check
/// actually consumes are required, so any superset of those fields
/// (the full `bench --json` output included) deserializes cleanly.
#[derive(Debug, Clone, Deserialize)]
pub struct BaselineRecord {
    pub operation: Operation,
    pub measured_p95_ms: f64,
}

/// Top-level shape of a `bench --json` payload, used to thread the
/// `results` array out for [`load_baseline`]. The other top-level
/// fields (`iterations`, `warmup`, anything future runs add) are
/// ignored on purpose so older / newer JSON shapes load without
/// migration churn.
#[derive(Debug, Clone, Deserialize)]
struct BaselineFile {
    results: Vec<BaselineRecord>,
}

/// Per-operation regression row produced by
/// [`compare_against_baseline`].
#[derive(Debug, Clone, Serialize)]
pub struct Regression {
    pub operation: Operation,
    /// Pretty label, duplicated for JSON consumers.
    pub label: &'static str,
    pub baseline_p95_ms: f64,
    pub measured_p95_ms: f64,
    pub delta_pct: f64,
    pub threshold_pct: f64,
    pub regressed: bool,
}

/// Load a previously emitted `bench --json` payload from disk.
///
/// # Errors
///
/// Returns an error if the file cannot be read or the JSON cannot be
/// parsed into the [`BaselineFile`] shape.
pub fn load_baseline(path: &Path) -> Result<Vec<BaselineRecord>> {
    let raw = std::fs::read_to_string(path)
        .with_context(|| format!("failed to read baseline file: {}", path.display()))?;
    let file: BaselineFile = serde_json::from_str(&raw)
        .with_context(|| format!("failed to parse baseline JSON: {}", path.display()))?;
    Ok(file.results)
}

/// Compare a fresh run against a baseline. Operations missing from the
/// baseline are skipped silently (e.g. a new bench row added since the
/// baseline was captured). The returned `Vec` preserves the order of
/// `current` and only includes ops present in both.
#[must_use]
pub fn compare_against_baseline(
    current: &[OperationResult],
    baseline: &[BaselineRecord],
    threshold_pct: f64,
) -> Vec<Regression> {
    let mut out = Vec::with_capacity(current.len());
    for r in current {
        let Some(b) = baseline.iter().find(|b| b.operation == r.operation) else {
            continue;
        };
        // Treat a non-positive baseline as "no signal" so we never
        // divide by zero or produce a nonsense -100% delta. Any current
        // measurement against a zero baseline is reported as 0% delta
        // rather than infinity — the absolute-budget guard already
        // catches actual breakage.
        let delta_pct = if b.measured_p95_ms > 0.0 {
            (r.measured_p95_ms - b.measured_p95_ms) / b.measured_p95_ms * 100.0
        } else {
            0.0
        };
        let regressed = delta_pct > threshold_pct;
        out.push(Regression {
            operation: r.operation,
            label: r.operation.label(),
            baseline_p95_ms: b.measured_p95_ms,
            measured_p95_ms: r.measured_p95_ms,
            delta_pct,
            threshold_pct,
            regressed,
        });
    }
    out
}

/// Render a regression table to a string, mirroring the layout of
/// [`render_table`].
#[must_use]
pub fn render_regression_table(rows: &[Regression]) -> String {
    let mut out = String::new();
    out.push_str(
        "Operation                       Baseline (p95)   Measured (p95)   Delta     Status\n",
    );
    out.push_str(
        "─────────────────────────────────────────────────────────────────────────────────\n",
    );
    for r in rows {
        let status_str = if r.regressed { "REGRESSION" } else { "OK" };
        let line = format!(
            "{:<30}  {:>10.1} ms     {:>10.1} ms    {:>+6.1}%   {}\n",
            r.label, r.baseline_p95_ms, r.measured_p95_ms, r.delta_pct, status_str
        );
        out.push_str(&line);
    }
    out
}

/// Append a benchmark result to a JSONL history file.
/// Creates the file and parent directories if missing.
/// Each line is a self-describing JSON object with `captured_at`, `iterations`,
/// `warmup`, and `results` array.
pub fn append_history(
    path: &std::path::Path,
    captured_at: &str,
    iterations: usize,
    warmup: usize,
    // #1579 B8 — corpus scale of the recorded run (`null` = default
    // workload) so downstream regression tooling can stratify history
    // entries per scale bucket.
    scale: Option<usize>,
    results: &[OperationResult],
) -> Result<()> {
    use std::fs::OpenOptions;
    use std::io::Write;

    // Create parent directories if needed
    if let Some(parent) = path.parent()
        && !parent.as_os_str().is_empty()
    {
        std::fs::create_dir_all(parent)?;
    }

    let entry = serde_json::json!({
        "captured_at": captured_at,
        "iterations": iterations,
        "warmup": warmup,
        "scale": scale,
        "results": results,
    });

    let mut file = OpenOptions::new().create(true).append(true).open(path)?;

    writeln!(file, "{}", serde_json::to_string(&entry)?)?;
    Ok(())
}

#[allow(clippy::wildcard_imports)]
mod tests {
    use super::*;
    use crate::db;

    #[allow(dead_code)]
    fn fresh_conn() -> Connection {
        db::open(Path::new(":memory:")).unwrap()
    }

    #[allow(dead_code)]
    fn small_config() -> BenchConfig {
        BenchConfig {
            iterations: 30,
            warmup: 5,
            namespace: "bench-test".to_string(),
            scale: None,
        }
    }

    #[test]
    fn percentile_interpolates() {
        let s = vec![1.0, 2.0, 3.0, 4.0];
        assert!((percentile(&s, 0.50) - 2.5).abs() < 1e-9);
        assert!((percentile(&s, 0.0) - 1.0).abs() < 1e-9);
        assert!((percentile(&s, 1.0) - 4.0).abs() < 1e-9);
    }

    #[test]
    fn percentile_handles_singleton_and_empty() {
        assert!((percentile(&[], 0.5) - 0.0).abs() < 1e-9);
        assert!((percentile(&[42.0], 0.99) - 42.0).abs() < 1e-9);
    }

    #[test]
    fn run_returns_all_seven_results() {
        let conn = fresh_conn();
        let results = run(&conn, &small_config()).unwrap();
        assert_eq!(results.len(), 7);
        assert_eq!(results[0].operation, Operation::StoreNoEmbedding);
        assert_eq!(results[1].operation, Operation::SearchFts);
        assert_eq!(results[2].operation, Operation::RecallHot);
        assert_eq!(results[3].operation, Operation::KgQueryDepth1);
        assert_eq!(results[4].operation, Operation::KgQueryDepth3);
        assert_eq!(results[5].operation, Operation::KgQueryDepth5);
        assert_eq!(results[6].operation, Operation::KgTimeline);
        for r in &results {
            assert_eq!(r.samples, 30);
            assert!(r.measured_p50_ms <= r.measured_p95_ms);
            assert!(r.measured_p95_ms <= r.measured_p99_ms);
            assert!(r.target_p95_ms > 0.0);
        }
    }

    #[test]
    fn status_is_fail_when_p95_over_tolerance() {
        let r = OperationResult {
            operation: Operation::StoreNoEmbedding,
            label: Operation::StoreNoEmbedding.label(),
            target_p95_ms: 20.0,
            measured_p50_ms: 5.0,
            measured_p95_ms: 25.0,
            measured_p99_ms: 30.0,
            samples: 100,
            status: Status::Fail,
        };
        assert_eq!(r.status, Status::Fail);
        // 25 > 20 * 1.10 = 22 → Fail
        let recomputed = if 25.0_f64 <= 20.0 * P95_TOLERANCE {
            Status::Pass
        } else {
            Status::Fail
        };
        assert_eq!(recomputed, Status::Fail);
    }

    #[test]
    fn status_is_pass_within_tolerance() {
        // 21 ms over 20 ms budget = 5% over → still PASS (under 10%).
        let recomputed = if 21.0_f64 <= 20.0 * P95_TOLERANCE {
            Status::Pass
        } else {
            Status::Fail
        };
        assert_eq!(recomputed, Status::Pass);
    }

    #[test]
    fn render_table_includes_all_operations() {
        let conn = fresh_conn();
        let results = run(&conn, &small_config()).unwrap();
        let table = render_table(&results);
        assert!(table.contains("memory_store (no embedding)"));
        assert!(table.contains("memory_search (FTS5)"));
        assert!(table.contains("memory_recall (hot, depth=1)"));
        assert!(table.contains("memory_kg_query (depth=1)"));
        assert!(table.contains("memory_kg_query (depth=3)"));
        assert!(table.contains("memory_kg_query (depth=5)"));
        assert!(table.contains("memory_kg_timeline"));
        assert!(table.contains("Status"));
    }

    #[test]
    fn operation_targets_match_performance_md() {
        // Pinned to PERFORMANCE.md — if you change a budget, change both.
        assert!((Operation::StoreNoEmbedding.target_p95_ms() - 20.0).abs() < 1e-9);
        assert!((Operation::SearchFts.target_p95_ms() - 100.0).abs() < 1e-9);
        assert!((Operation::RecallHot.target_p95_ms() - 50.0).abs() < 1e-9);
        assert!((Operation::KgQueryDepth1.target_p95_ms() - 100.0).abs() < 1e-9);
        assert!((Operation::KgQueryDepth3.target_p95_ms() - 100.0).abs() < 1e-9);
        assert!((Operation::KgQueryDepth5.target_p95_ms() - 250.0).abs() < 1e-9);
        assert!((Operation::KgTimeline.target_p95_ms() - 100.0).abs() < 1e-9);
    }

    /// Issue #1193 — the effective budget the pass/fail verdict uses
    /// is the canonical budget × `MACOS_BUDGET_MULT`. On Linux/Windows
    /// the multiplier is 1.0 (effective == canonical); on macOS the
    /// multiplier is 3.0 (effective == 3 × canonical). Regression-pins
    /// the wiring so a future refactor can't silently revert the
    /// platform-aware verdict path.
    #[test]
    fn effective_target_applies_macos_multiplier() {
        for op in [
            Operation::StoreNoEmbedding,
            Operation::SearchFts,
            Operation::RecallHot,
            Operation::KgQueryDepth1,
            Operation::KgQueryDepth3,
            Operation::KgQueryDepth5,
            Operation::KgTimeline,
        ] {
            let expected = op.target_p95_ms() * MACOS_BUDGET_MULT;
            assert!(
                (op.effective_target_p95_ms() - expected).abs() < 1e-9,
                "effective budget for {:?} = {} (expected {})",
                op,
                op.effective_target_p95_ms(),
                expected,
            );
        }
        #[cfg(target_os = "macos")]
        assert!((MACOS_BUDGET_MULT - 3.0).abs() < 1e-9);
        #[cfg(not(target_os = "macos"))]
        assert!((MACOS_BUDGET_MULT - 1.0).abs() < 1e-9);
    }

    /// #1579 B8 — pins the per-scale budget table to the values
    /// published in `PERFORMANCE.md` §"Corpus-scale budgets". If you
    /// change a scale budget, change both.
    #[test]
    fn operation_scale_targets_match_performance_md() {
        let at_gate_scale = Some(CI_SCALE_GATE_ROWS);
        assert!(
            (Operation::StoreNoEmbedding.target_p95_ms_at_scale(at_gate_scale) - 120.0).abs()
                < 1e-9
        );
        assert!((Operation::SearchFts.target_p95_ms_at_scale(at_gate_scale) - 60.0).abs() < 1e-9);
        assert!((Operation::RecallHot.target_p95_ms_at_scale(at_gate_scale) - 80.0).abs() < 1e-9);
        // KG fixtures are scale-independent → canonical budgets hold.
        for op in [
            Operation::KgQueryDepth1,
            Operation::KgQueryDepth3,
            Operation::KgQueryDepth5,
            Operation::KgTimeline,
        ] {
            assert!(
                (op.target_p95_ms_at_scale(at_gate_scale) - op.target_p95_ms()).abs() < 1e-9,
                "{op:?} must keep its canonical budget at scale"
            );
        }
        // `None` (default workload) keeps the legacy budgets.
        assert!((Operation::RecallHot.target_p95_ms_at_scale(None) - 50.0).abs() < 1e-9);
    }

    /// #1579 B8 — bucket resolution: a request at or below a pinned
    /// scale selects that row; a request beyond the largest pinned
    /// scale falls back to the largest row (best-effort).
    #[test]
    fn issue_1579_b8_scale_budget_bucket_resolution() {
        assert_eq!(scale_budgets_for(500).scale, CI_SCALE_GATE_ROWS);
        assert_eq!(
            scale_budgets_for(CI_SCALE_GATE_ROWS).scale,
            CI_SCALE_GATE_ROWS
        );
        assert_eq!(scale_budgets_for(MAX_SCALE).scale, CI_SCALE_GATE_ROWS);
    }

    /// #1579 B8 — a `--scale` run actually seeds the scratch corpus
    /// (the P1 failure mode was a bench that never grew the table) and
    /// gates the three corpus-sensitive ops against the scale budgets.
    #[test]
    fn issue_1579_b8_scale_run_seeds_corpus_and_uses_scale_budgets() {
        let conn = fresh_conn();
        let ns = "bench-scale-test";
        let config = BenchConfig {
            iterations: 10,
            warmup: 2,
            namespace: ns.to_string(),
            scale: Some(300),
        };
        let results = run(&conn, &config).unwrap();
        assert_eq!(results.len(), 7);
        let seeded: i64 = conn
            .query_row(
                "SELECT COUNT(*) FROM memories WHERE namespace = ?1",
                [ns],
                |r| r.get(0),
            )
            .unwrap();
        assert!(
            seeded >= 300,
            "scale run must seed the scratch corpus; found {seeded} rows"
        );
        // Scale budgets reported (300 resolves into the 10k bucket).
        let store = &results[0];
        assert_eq!(store.operation, Operation::StoreNoEmbedding);
        assert!((store.target_p95_ms - 120.0).abs() < 1e-9);
        let search = &results[1];
        assert!((search.target_p95_ms - 60.0).abs() < 1e-9);
        let recall = &results[2];
        assert!((recall.target_p95_ms - 80.0).abs() < 1e-9);
        // KG rows keep canonical budgets.
        assert!((results[3].target_p95_ms - 100.0).abs() < 1e-9);
        assert!((results[5].target_p95_ms - 250.0).abs() < 1e-9);
    }

    #[test]
    fn seed_kg_chain_fixture_traverses_to_max_depth() {
        let conn = fresh_conn();
        let sources = seed_kg_chain_fixture(&conn, "kg-chain-fixture-test").unwrap();
        assert_eq!(sources.len(), KG_CHAIN_FIXTURE_CHAINS);
        // Every chain must yield exactly `KG_CHAIN_FIXTURE_HOPS` reachable
        // nodes at depth=KG_CHAIN_FIXTURE_HOPS — that's what justifies the
        // depth=5 budget bucket. depth=3 must reach exactly 3 nodes.
        for src in &sources {
            let depth5 =
                db::kg_query(&conn, src, KG_CHAIN_FIXTURE_HOPS, None, None, None, false).unwrap();
            assert_eq!(
                depth5.len(),
                KG_CHAIN_FIXTURE_HOPS,
                "depth={KG_CHAIN_FIXTURE_HOPS} on a {KG_CHAIN_FIXTURE_HOPS}-hop chain must reach every node"
            );
            let depth3 = db::kg_query(&conn, src, 3, None, None, None, false).unwrap();
            assert_eq!(
                depth3.len(),
                3,
                "depth=3 on a {KG_CHAIN_FIXTURE_HOPS}-hop chain must reach exactly 3 follow-on nodes"
            );
        }
    }

    #[test]
    fn seed_kg_fixture_populates_sources_and_links() {
        let conn = fresh_conn();
        let sources = seed_kg_fixture(&conn, "kg-fixture-test").unwrap();
        assert_eq!(sources.len(), KG_FIXTURE_SOURCES);
        // Every source carries the expected fan-out, every link has a
        // non-null `valid_from` (otherwise `kg_timeline` would skip it).
        for src in &sources {
            let nodes = db::kg_query(&conn, src, 1, None, None, None, false).unwrap();
            assert_eq!(nodes.len(), KG_FIXTURE_LINKS_PER_SOURCE);
            let timeline = db::kg_timeline(&conn, src, None, None, None).unwrap();
            assert_eq!(timeline.len(), KG_FIXTURE_LINKS_PER_SOURCE);
            for ev in &timeline {
                // `kg_timeline` filters out NULL `valid_from` rows in SQL,
                // so any returned event must carry a non-empty stamp.
                assert!(
                    !ev.valid_from.is_empty(),
                    "kg fixture must stamp valid_from on every link"
                );
            }
        }
    }

    #[allow(dead_code)]
    fn synthetic_result(op: Operation, p95: f64) -> OperationResult {
        OperationResult {
            operation: op,
            label: op.label(),
            target_p95_ms: op.target_p95_ms(),
            measured_p50_ms: p95 / 2.0,
            measured_p95_ms: p95,
            measured_p99_ms: p95 * 1.1,
            samples: 100,
            status: Status::Pass,
        }
    }

    #[allow(dead_code)]
    fn synthetic_baseline(op: Operation, p95: f64) -> BaselineRecord {
        BaselineRecord {
            operation: op,
            measured_p95_ms: p95,
        }
    }

    #[test]
    fn baseline_compare_flags_above_threshold() {
        // 12% slowdown over baseline at default 10% threshold → REGRESSION.
        let current = vec![synthetic_result(Operation::StoreNoEmbedding, 11.2)];
        let baseline = vec![synthetic_baseline(Operation::StoreNoEmbedding, 10.0)];
        let rows = compare_against_baseline(&current, &baseline, 10.0);
        assert_eq!(rows.len(), 1);
        assert!(rows[0].regressed);
        assert!((rows[0].delta_pct - 12.0).abs() < 1e-9);
    }

    #[test]
    fn baseline_compare_passes_within_threshold() {
        // 8% slowdown over baseline at default 10% threshold → OK.
        let current = vec![synthetic_result(Operation::StoreNoEmbedding, 10.8)];
        let baseline = vec![synthetic_baseline(Operation::StoreNoEmbedding, 10.0)];
        let rows = compare_against_baseline(&current, &baseline, 10.0);
        assert_eq!(rows.len(), 1);
        assert!(!rows[0].regressed);
    }

    #[test]
    fn baseline_compare_speedup_is_negative_delta() {
        // Faster than baseline → negative delta, never a regression.
        let current = vec![synthetic_result(Operation::SearchFts, 8.0)];
        let baseline = vec![synthetic_baseline(Operation::SearchFts, 10.0)];
        let rows = compare_against_baseline(&current, &baseline, 10.0);
        assert_eq!(rows.len(), 1);
        assert!(!rows[0].regressed);
        assert!((rows[0].delta_pct + 20.0).abs() < 1e-9);
    }

    #[test]
    fn baseline_compare_skips_ops_missing_in_baseline() {
        // A new op added since the baseline was captured shouldn't crash
        // or appear as a regression.
        let current = vec![
            synthetic_result(Operation::StoreNoEmbedding, 10.0),
            synthetic_result(Operation::KgQueryDepth5, 200.0),
        ];
        let baseline = vec![synthetic_baseline(Operation::StoreNoEmbedding, 10.0)];
        let rows = compare_against_baseline(&current, &baseline, 10.0);
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].operation, Operation::StoreNoEmbedding);
    }

    #[test]
    fn baseline_compare_handles_zero_baseline() {
        // Pathological zero baseline: report 0% delta rather than
        // dividing by zero. Absolute-budget guard still catches
        // genuinely-broken measurements.
        let current = vec![synthetic_result(Operation::SearchFts, 5.0)];
        let baseline = vec![synthetic_baseline(Operation::SearchFts, 0.0)];
        let rows = compare_against_baseline(&current, &baseline, 10.0);
        assert_eq!(rows.len(), 1);
        assert!(!rows[0].regressed);
        assert!((rows[0].delta_pct - 0.0).abs() < 1e-9);
    }

    #[test]
    fn load_baseline_round_trips_json_payload() {
        // Mirror the shape `bench --json` actually emits — it must
        // round-trip through `load_baseline` so CI artifacts work as
        // baselines without preprocessing.
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("baseline.json");
        let payload = serde_json::json!({
            "iterations": 200,
            "warmup": 20,
            "results": [
                {
                    "operation": "store_no_embedding",
                    "label": "memory_store (no embedding)",
                    "target_p95_ms": 20.0,
                    "measured_p50_ms": 4.0,
                    "measured_p95_ms": 9.0,
                    "measured_p99_ms": 11.0,
                    "samples": 200,
                    "status": "pass"
                },
                {
                    "operation": "search_fts",
                    "label": "memory_search (FTS5)",
                    "target_p95_ms": 100.0,
                    "measured_p50_ms": 12.0,
                    "measured_p95_ms": 31.0,
                    "measured_p99_ms": 45.0,
                    "samples": 200,
                    "status": "pass"
                }
            ]
        });
        std::fs::write(&path, serde_json::to_string_pretty(&payload).unwrap()).unwrap();
        let loaded = load_baseline(&path).unwrap();
        assert_eq!(loaded.len(), 2);
        assert_eq!(loaded[0].operation, Operation::StoreNoEmbedding);
        assert!((loaded[0].measured_p95_ms - 9.0).abs() < 1e-9);
        assert_eq!(loaded[1].operation, Operation::SearchFts);
        assert!((loaded[1].measured_p95_ms - 31.0).abs() < 1e-9);
    }

    #[test]
    fn render_regression_table_marks_regressions() {
        let rows = vec![
            Regression {
                operation: Operation::StoreNoEmbedding,
                label: Operation::StoreNoEmbedding.label(),
                baseline_p95_ms: 10.0,
                measured_p95_ms: 12.0,
                delta_pct: 20.0,
                threshold_pct: 10.0,
                regressed: true,
            },
            Regression {
                operation: Operation::SearchFts,
                label: Operation::SearchFts.label(),
                baseline_p95_ms: 30.0,
                measured_p95_ms: 31.0,
                delta_pct: 3.3,
                threshold_pct: 10.0,
                regressed: false,
            },
        ];
        let table = render_regression_table(&rows);
        assert!(table.contains("memory_store (no embedding)"));
        assert!(table.contains("memory_search (FTS5)"));
        assert!(table.contains("REGRESSION"));
        assert!(table.contains("OK"));
    }
}