areev_loop/

recommendation.rs

//! The recommendation object (OMS 0x0C), the deterministic summary renderer,
//! the dedup key, and the lifecycle state machine + audit records.
//!
//! Invariants enforced here:
//! - Analyzers emit a `(template_id, args)` summary, never free prose.
//! - `dedup_key`, `origin`, and the params snapshot are engine-stamped.
//! - `dedup_key` excludes proposal content and the `/major` version, so a
//!   growing cluster or an analyzer upgrade does not re-propose as novel —
//!   for ANALYZER findings. An authored (`origin = llm`) executable proposal
//!   keys on a fingerprint of its content as well, because there the content
//!   IS the finding: two different lessons on one entity are two findings,
//!   and the same lesson re-authored is one.
//! - Lifecycle transitions are gated; `pending → applied` is policy-only.

use crate::error::{Error, Result};
use crate::model::{normalize_ident, ActionKind, Origin, Severity};
use crate::substrate::GrainSpec;
use serde::{Deserialize, Serialize};
use serde_json::{Map, Value};

/// A deterministic, template-rendered summary. The analyzer chooses a
/// `template_id` and supplies `args`; the text is produced here, so an
/// analyzer can never emit arbitrary prose into the queue.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct Summary {
    pub template_id: String,
    #[serde(default)]
    pub args: Map<String, Value>,
}

impl Summary {
    pub fn new(template_id: impl Into<String>, args: Map<String, Value>) -> Self {
        Summary {
            template_id: template_id.into(),
            args,
        }
    }

    /// Render the summary. Unknown template ids fall back to a stable, honest
    /// string (never a panic) so a manifest/template mismatch is visible, not
    /// fatal.
    pub fn render(&self) -> String {
        let t = builtin_template(&self.template_id).unwrap_or("{template_id}: {summary}");
        interpolate(t, &self.args, &self.template_id)
    }
}

/// The built-in template table. Deterministic; the only place summary prose
/// lives. Keep placeholders in `{name}` form matching `args` keys.
fn builtin_template(id: &str) -> Option<&'static str> {
    Some(match id {
        "duplicate.exact" => "Consolidate {count} exact-duplicate grains for \"{subject}\"",
        "duplicate.near" => {
            "Consolidate {count} near-duplicate observations (similarity ≥ {threshold})"
        }
        "contradiction.functional" => {
            "\"{subject}\" holds {count} live values for functional relation \"{relation}\""
        }
        "tool_failure.cluster" => {
            "Tool \"{tool}\" failed {count} times ({rate}% of the calls that could \
             fail this way): {signature}"
        }
        "staleness.expired" => "Expire \"{subject}\": past its declared valid_to ({age_days}d ago)",
        "fork.multi_head" => "Entity \"{entity}\" has {count} competing heads",
        "skill.stall" => {
            "Skill \"{skill}\" isn't improving: practiced {practice_count}× but proficiency is still {proficiency}"
        }
        "goal.stagnation" => "Goal \"{goal}\" is stalled: active {age_days}d with {progress} progress",
        "cold.grain" => {
            "Cold memory: \"{subject}\" ({age_days}d old) has never been recalled — retire candidate"
        }
        "coverage.gap" => {
            "Recurring question with no matching memory: \"{query}\" asked {count}× ({empty_rate}% empty)"
        }
        "budget.pressure" => {
            "Assembly budget overflowed on {overflow_rate}% of {samples} recalls — raise the budget or curate memory"
        }
        "retention.overdue" => {
            "Retention: {count} grain(s) in \"{namespace}\" exceed the {max_age_days}d policy (oldest {oldest_days}d); {remaining} more await a later pass"
        }
        "outcome.regression" => {
            "Applied recommendation regressed: {metric} moved {baseline} → {current}"
        }
        // The high-water baseline names both figures — the best run before
        // the apply and the run that fell from it — because the reviewer has
        // to judge whether THIS rule caused the whole fall from the peak.
        "outcome.regression_high_water" => {
            "Applied recommendation regressed against the best run before the apply ({baseline_run}): {metric} moved {baseline} → {current}"
        }
        // A regression that also breached the cost bound: the revert names
        // the cost delta so the reviewer sees both halves of the damage.
        "outcome.regression_costlier" => {
            "Applied recommendation regressed: {metric} moved {baseline} → {current}, and {cost_field} rose {cost_baseline} → {cost_current} (bound ×{cost_ratio})"
        }
        "outcome.regression_high_water_costlier" => {
            "Applied recommendation regressed against the best run before the apply ({baseline_run}): {metric} moved {baseline} → {current}, and {cost_field} rose {cost_baseline} → {cost_current} (bound ×{cost_ratio})"
        }
        // Quality held, cost did not: advisory only — a cost/quality trade
        // is a human decision, so this is a Flag, never a revert draft.
        "outcome.held_costlier" => {
            "Applied recommendation held ({metric} {baseline} → {current}) but {cost_field} rose {cost_baseline} → {cost_current}, past ×{cost_ratio} of the baseline run ({baseline_run} → {current_run}) — a cost/quality trade to decide"
        }
        "outcome.premise_drift" => {
            "{current} of the grains this recommendation cited have since been superseded by a different value or retracted — its premise moved; revert it"
        }
        "run.failures" => {
            "Workflow {workflow} failed {failed}/{runs} recent runs ({rate}%): {last_error}"
        }
        "run.cost" => {
            "Workflow {workflow} spent ${usd} across {runs} runs (avg ${avg_usd}/run)"
        }
        "run.context_pressure" => {
            "Workflow {workflow} summarized its own transcript {folds} time(s) across {runs} runs (avg {avg_folds}/run) — its nodes are outgrowing the model's window"
        }
        "adapter.candidate" => {
            "Promote adapter for \"{model}\" (base {base_model}) — gated on its pinned evalset"
        }
        // origin=llm drafts carry free text (clearly marked llm) rather than a
        // deterministic template — the model's proposed summary rides in {text}.
        "llm.discover" => "{text}",
        // An LLM-authored lesson shows the reviewer BOTH the finding and the
        // exact line an apply would record — never one without the other.
        "llm.lesson" => "{text} — record lesson: \"{lesson}\"",
        // The same lesson, flagged as restating a live one: the reviewer sees
        // that a rule of this meaning already exists before approving a
        // second copy of it.
        "llm.lesson_near_duplicate" => {
            "{text} — record lesson: \"{lesson}\" — NEAR-DUPLICATE of {near_count} live lesson(s) on this entity (best match {near_score} by {near_method}, {near_hash})"
        }
        // A consolidation: one lesson replacing a pile. The apply supersedes
        // every member and adds the one line; rollback restores them all.
        "llm.consolidation" => {
            "{text} — consolidate {count} lessons into one: \"{lesson}\""
        }
        "lesson.pile" => {
            "{count} live lessons on \"{subject}\" exceed the budget of {max_active} — every rule competes for the model's attention; consolidate or retire: {members}"
        }
        // The rest of the LLM proposal vocabulary follows the same rule: the
        // finding AND the exact change an apply would make, never one without
        // the other. A reviewer approving blind is the failure this prevents.
        "llm.fact" => "{text} — record fact: {relation} = \"{object}\"",
        "llm.skill" => "{text} — record skill: \"{name}\" ({steps} steps)",
        "llm.plan" => "{text} — record plan: \"{name}\" ({nodes} steps, {edges} edges)",
        "llm.query_revision" => "{text} — redefine \"{name}\" as: {body}",
        "llm.plan_revision" => "{text} — revise plan {plan}: {edits}",
        // The rehearsal refused it: the edit is shown, the reason names the
        // runs, and nothing is offered to apply.
        "llm.plan_revision_refused" => {
            "{text} — plan revision of {plan} ({edits}) NOT applicable: rehearsed against the plan's journaled runs and {reason}"
        }
        "llm.code_revision" => {
            "{text} — new source for tool \"{tool}\" ({bytes} bytes), pending its evalset gate"
        }
        // External command analyzers (trust class Command): free text from the
        // subprocess, rendered as-is (the trust badge, not the prose, marks it).
        "command.finding" => "{text}",
        _ => return None,
    })
}

fn interpolate(template: &str, args: &Map<String, Value>, template_id: &str) -> String {
    let mut out = String::with_capacity(template.len());
    let mut chars = template.chars().peekable();
    while let Some(c) = chars.next() {
        if c == '{' {
            let mut key = String::new();
            for k in chars.by_ref() {
                if k == '}' {
                    break;
                }
                key.push(k);
            }
            if key == "template_id" {
                out.push_str(template_id);
            } else {
                match args.get(&key) {
                    Some(Value::String(s)) => out.push_str(s),
                    Some(v) => out.push_str(&v.to_string()),
                    None => {
                        // Missing arg — leave a visible marker rather than lie.
                        out.push('{');
                        out.push_str(&key);
                        out.push('}');
                    }
                }
            }
        } else {
            out.push(c);
        }
    }
    out
}

/// When an applied recommendation is re-measured — one checkpoint of the
/// Verify gate's schedule, in the unit the deployment actually counts in.
///
/// Three units, because deployments count differently and a fixed one makes
/// the gate inert everywhere else. A service evaluated nightly counts
/// **time** (`after_ms`). A benchmark or a CI harness counts **graded runs**
/// (`after_runs`): "measure at the next evaluation after the apply", however
/// long that takes on the clock. A chat agent counts **turns** (`after_grains`):
/// "after fifty more grains". The default schedule stayed ms-only for a year
/// and was measured, on PAST-Bench, to fire exactly zero verdicts across 78
/// governed runs — a family finishes in seven minutes and the first checkpoint
/// was a day away (`crates/areev-bench/PERSIST.md`).
///
/// On the wire a bare integer is milliseconds, so every policy, metric
/// snapshot and state blob written before this type existed reads back
/// unchanged, and an all-ms schedule still serializes as `[86400000, …]`.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub enum Checkpoint {
    /// Elapsed time since the apply.
    AfterMs(i64),
    /// Evalset runs journaled since the apply. Only an `evalset:` metric can
    /// count these; on any other metric the checkpoint never comes due.
    AfterRuns(u32),
    /// Grains written since the apply — the loop's own notion of "activity".
    AfterGrains(u32),
}

impl Checkpoint {
    /// The label a person reads: `1d`, `12h`, `3 runs`, `50 grains`.
    pub fn label(&self) -> String {
        match self {
            Checkpoint::AfterMs(ms) if ms % 86_400_000 == 0 => format!("{}d", ms / 86_400_000),
            Checkpoint::AfterMs(ms) if ms % 3_600_000 == 0 => format!("{}h", ms / 3_600_000),
            Checkpoint::AfterMs(ms) => format!("{ms}ms"),
            Checkpoint::AfterRuns(n) => format!("{n} run{}", if *n == 1 { "" } else { "s" }),
            Checkpoint::AfterGrains(n) => format!("{n} grain{}", if *n == 1 { "" } else { "s" }),
        }
    }

    /// The ms value when this is a time checkpoint — what legacy consumers of
    /// `OutcomeResult::horizon_ms` read.
    pub fn as_ms(&self) -> Option<i64> {
        match self {
            Checkpoint::AfterMs(ms) => Some(*ms),
            _ => None,
        }
    }
}

impl Serialize for Checkpoint {
    fn serialize<S: serde::Serializer>(&self, ser: S) -> std::result::Result<S::Ok, S::Error> {
        use serde::ser::SerializeMap;
        match self {
            // Bare integer: byte-identical to the pre-Checkpoint `Vec<i64>`.
            Checkpoint::AfterMs(ms) => ser.serialize_i64(*ms),
            Checkpoint::AfterRuns(n) => {
                let mut m = ser.serialize_map(Some(1))?;
                m.serialize_entry("after_runs", n)?;
                m.end()
            }
            Checkpoint::AfterGrains(n) => {
                let mut m = ser.serialize_map(Some(1))?;
                m.serialize_entry("after_grains", n)?;
                m.end()
            }
        }
    }
}

impl<'de> Deserialize<'de> for Checkpoint {
    fn deserialize<D: serde::Deserializer<'de>>(de: D) -> std::result::Result<Self, D::Error> {
        use serde::de::Error as _;
        let v = Value::deserialize(de)?;
        match &v {
            Value::Number(n) => n
                .as_i64()
                .filter(|ms| *ms >= 0)
                .map(Checkpoint::AfterMs)
                .ok_or_else(|| D::Error::custom("checkpoint: a bare number is non-negative milliseconds")),
            Value::Object(m) if m.len() == 1 => {
                let (k, val) = m.iter().next().unwrap();
                let n = val.as_u64().ok_or_else(|| {
                    D::Error::custom(format!("checkpoint: {k} takes a non-negative integer"))
                })?;
                match k.as_str() {
                    "after_ms" => i64::try_from(n)
                        .map(Checkpoint::AfterMs)
                        .map_err(|_| D::Error::custom("checkpoint: after_ms out of range")),
                    "after_runs" => u32::try_from(n)
                        .map(Checkpoint::AfterRuns)
                        .map_err(|_| D::Error::custom("checkpoint: after_runs out of range")),
                    "after_grains" => u32::try_from(n)
                        .map(Checkpoint::AfterGrains)
                        .map_err(|_| D::Error::custom("checkpoint: after_grains out of range")),
                    other => Err(D::Error::custom(format!(
                        "checkpoint: unknown unit {other:?} (expected after_ms, after_runs or after_grains)"
                    ))),
                }
            }
            _ => Err(D::Error::custom(
                "checkpoint: expected milliseconds or {\"after_ms\"|\"after_runs\"|\"after_grains\": n}",
            )),
        }
    }
}

/// A reproducible metric snapshot; powers outcome review.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct MetricSnapshot {
    /// Metric kind — the engine knows how to re-measure a fixed set
    /// (e.g. `tool_error_recurrence`); unknown kinds are skipped, not faked.
    pub metric: String,
    pub baseline: f64,
    pub unit: String,
    pub n: u64,
    pub window: String,
    /// The subject the metric is about (e.g. the tool name), used by the
    /// engine's typed re-measurement.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub subject: Option<String>,
    /// Namespace scope for the re-measurement, normalized (case-fold/trim).
    /// `None` = all namespaces. Additive: older snapshots deserialize to
    /// `None` and existing metric kinds ignore it.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub namespace: Option<String>,
    /// Relation scope for fact-shaped metrics (e.g. which functional relation
    /// a contradiction was resolved under), normalized. Additive like
    /// `namespace`.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub relation: Option<String>,
    /// CAL that recomputes the metric at verify time (reproducibility /
    /// documentation; the engine re-measures with typed reads).
    pub query: String,
    /// How long after apply to re-measure, in epoch-ms delta (the first / only
    /// checkpoint when `horizons_ms` is empty).
    pub review_after_ms: i64,
    /// A schedule of checkpoints (ms after apply) to re-measure at. An outcome
    /// that `held` at an early checkpoint can `regress` at a later one, so a
    /// verdict is never final until the last horizon. Empty → `[review_after_ms]`.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub horizons_ms: Vec<i64>,
    /// The schedule in the deployment's own unit ([`Checkpoint`]). When set it
    /// is THE schedule and `horizons_ms`/`review_after_ms` are ignored; empty
    /// (every snapshot written before it existed) falls back to them.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub checkpoints: Vec<Checkpoint>,
    /// Which direction is an improvement. The built-in metrics are all
    /// recurrence counts, where lower is better — so this defaults to `false`
    /// and every existing snapshot deserializes unchanged. An evalset accuracy
    /// is the opposite, and getting it wrong does not merely misreport: the
    /// Verify gate would read a rule that IMPROVED accuracy as a regression and
    /// propose reverting it.
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub higher_is_better: bool,
}

/// The one regression rule.
///
/// The verdict is computed in three places — by the engine for the recorded
/// `OutcomeResult`, by `outcome_review` when it drafts the revert, and by
/// `areev eval run --tolerance` when it re-accepts a model swap — and any two
/// silently disagreeing would either revert a change that held or sit on one
/// that regressed. So all of them call this.
///
/// `tolerance` is the policy's minimum effect size in the metric's own unit
/// (`OutcomeEvalset::min_effect`, or `--tolerance` in percentage points of
/// the pass rate): a worsening of at most that much is not a regression. It
/// is a **floor, not a significance test** — no p-value, no interval; a host
/// that needs statistics has the run counts to compute them. Zero is the
/// pre-policy behaviour: any drop is a regression. `EPSILON` is only
/// floating-point equality slack, so a value read twice that differs in the
/// last bit does not flip the verdict; it is not a noise floor.
pub fn is_regression(baseline: f64, current: f64, higher_is_better: bool, tolerance: f64) -> bool {
    const EPSILON: f64 = 1e-9;
    let tolerance = if tolerance.is_finite() && tolerance > 0.0 { tolerance } else { 0.0 };
    if higher_is_better {
        current < baseline - tolerance - EPSILON
    } else {
        current > baseline + tolerance + EPSILON
    }
}

impl MetricSnapshot {
    /// The measurement schedule, sorted and deduplicated: `checkpoints` when
    /// set, else `horizons_ms`, else the single `review_after_ms` — each older
    /// spelling read as time, so a snapshot from before [`Checkpoint`] existed
    /// measures exactly as it did.
    pub fn schedule(&self) -> Vec<Checkpoint> {
        let mut h: Vec<Checkpoint> = if !self.checkpoints.is_empty() {
            self.checkpoints.clone()
        } else if !self.horizons_ms.is_empty() {
            self.horizons_ms.iter().map(|ms| Checkpoint::AfterMs(*ms)).collect()
        } else {
            vec![Checkpoint::AfterMs(self.review_after_ms)]
        };
        h.sort_unstable();
        h.dedup();
        h
    }

    /// The time checkpoints of the schedule, in ms — the pre-[`Checkpoint`]
    /// view, kept for callers that only understand time.
    pub fn horizons(&self) -> Vec<i64> {
        self.schedule().iter().filter_map(Checkpoint::as_ms).collect()
    }
}

fn is_zero(x: &f64) -> bool {
    *x == 0.0
}

/// What a checkpoint read of the policy's cost bound, beside the quality
/// verdict. `status` is `within`, `breached`, or `not_measurable` (the
/// field was absent or malformed on either run — then `baseline`/`current`
/// carry whichever side did measure, or nothing). Present on a record only
/// when the policy set a bound.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct CostRead {
    pub field: String,
    pub max_increase_ratio: f64,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub baseline: Option<f64>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub current: Option<f64>,
    pub status: String,
}

impl CostRead {
    pub fn breached(&self) -> bool {
        self.status == "breached"
    }
}

/// A measured outcome for an applied recommendation at one checkpoint — the
/// Verify gate's output. `held` = the metric did not regress at this horizon;
/// `regressed` = it got worse (a revert is proposed). A recommendation
/// accumulates one of these per horizon, forming a time series.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct OutcomeResult {
    pub rec_hash: String,
    pub metric: String,
    pub baseline: f64,
    pub current: f64,
    pub verdict: String,
    /// Where `baseline` came from, so the receipt names the run it compared
    /// against: `newest_before_apply` or `high_water` (an evalset run, named
    /// in `baseline_run_id`), or `snapshot` — the number the proposal froze,
    /// used when no run was journaled before the apply or the metric is not
    /// evalset-backed. Empty on a record written before this field existed.
    #[serde(default, skip_serializing_if = "String::is_empty")]
    pub baseline_kind: String,
    /// The evalset run `baseline` was read from, when it was read from one.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub baseline_run_id: Option<String>,
    /// Advisory, independent of the policy's baseline choice: the best value
    /// the field attained on any run journaled before the apply. When the
    /// verdict is `held` against a newer, lower baseline this is the lost
    /// opportunity the marginal comparison cannot see.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub best_before: Option<f64>,
    /// The minimum effect size the verdict was made under, in the metric's
    /// unit (`OutcomeEvalset::min_effect`, resolved at measure time). Absent
    /// when zero, so a `held` under a floor is distinguishable from a `held`
    /// at zero and a record without one reads as before.
    #[serde(default, skip_serializing_if = "is_zero")]
    pub tolerance: f64,
    /// The evalset run `current` was read from, when it was read from one.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub current_run_id: Option<String>,
    /// The cost bound's reading at this checkpoint, when the policy set one.
    /// A quality `held` with a breached bound records the verdict
    /// `held_costlier`; `regressed` dominates.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cost: Option<CostRead>,
    /// Which checkpoint this measurement is for (ms after apply). Zero for a
    /// checkpoint counted in runs or grains — see `checkpoint`.
    #[serde(default)]
    pub horizon_ms: i64,
    /// The checkpoint in its own unit, when that unit is not time. A time
    /// checkpoint leaves this unset and speaks through `horizon_ms`, so every
    /// consumer of the older shape reads an unchanged record.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub checkpoint: Option<Checkpoint>,
    pub measured_at_ms: i64,
}

/// The proposed change. Exactly one variant per recommendation.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
#[serde(tag = "proposal", rename_all = "snake_case")]
pub enum Proposal {
    /// A batch of CAL Tier-1 evolve writes (ADD/SUPERSEDE), MAY contain FORGET.
    Cal { cal: String },
    /// A doc-target edit: `{format, base_digest, diff}` (base_digest enables a
    /// staleness check at apply).
    Edit {
        format: String,
        base_digest: String,
        diff: String,
    },
    /// An opaque map for host targets (applied by the host, §12.3).
    Data { data: Map<String, Value> },
}

/// What an analyzer emits. `dedup_key`, `origin`, and the params snapshot are
/// **not** here — the engine stamps them. Non-exhaustive so the engine can add
/// fields without breaking analyzers.
#[derive(Debug, Clone, PartialEq)]
#[non_exhaustive]
pub struct RecDraft {
    pub target_ref: String,
    pub action_kind: ActionKind,
    pub summary: Summary,
    pub severity: Severity,
    pub proposal: Proposal,
    /// Bounded to ≤64 representative evidence hashes by the engine.
    pub evidence: Vec<String>,
    pub evidence_query: Option<String>,
    pub metric: Option<MetricSnapshot>,
    pub confidence: f64,
    pub importance: f64,
    /// Rule E1 pin for `code_revision` drafts (see [`validate_code_rules`]).
    pub evalset_hash: Option<String>,
}

impl RecDraft {
    pub fn new(
        target_ref: impl Into<String>,
        action_kind: ActionKind,
        summary: Summary,
        proposal: Proposal,
    ) -> Self {
        RecDraft {
            target_ref: target_ref.into(),
            action_kind,
            summary,
            severity: Severity::Low,
            proposal,
            evidence: Vec::new(),
            evidence_query: None,
            metric: None,
            confidence: 0.8,
            importance: 0.5,
            evalset_hash: None,
        }
    }

    pub fn severity(mut self, s: Severity) -> Self {
        self.severity = s;
        self
    }
    pub fn evidence(mut self, hashes: Vec<String>) -> Self {
        self.evidence = hashes;
        self
    }
    pub fn evidence_query(mut self, q: impl Into<String>) -> Self {
        self.evidence_query = Some(q.into());
        self
    }
    pub fn metric(mut self, m: MetricSnapshot) -> Self {
        self.metric = Some(m);
        self
    }
    pub fn confidence(mut self, c: f64) -> Self {
        self.confidence = c;
        self
    }
    pub fn importance(mut self, i: f64) -> Self {
        self.importance = i;
        self
    }
    pub fn evalset_hash(mut self, h: impl Into<String>) -> Self {
        self.evalset_hash = Some(h.into());
        self
    }
}

/// Maximum representative evidence hashes carried inline (proposal §7.1).
pub const MAX_EVIDENCE: usize = 64;

/// Compute the dedup key: `family ⟂ target_ref ⟂ action_kind`, case-folded.
/// Excludes proposal content and evidence by construction.
pub fn dedup_key(family: &str, target_ref: &str, action: ActionKind) -> String {
    // U+001F (unit separator) cannot appear in any of the inputs.
    format!(
        "{}\u{1f}{}\u{1f}{}",
        normalize_ident(family),
        normalize_ident(target_ref),
        action.as_str()
    )
}

/// The dedup key of an AUTHORED executable proposal: [`dedup_key`] plus a
/// fingerprint of the proposal's content. An analyzer finding is "this
/// target has this kind of problem", so content is rightly excluded; an
/// authored lesson is "do this", and two different lessons on the same
/// entity must both reach the queue while the same lesson re-authored must
/// not. The fingerprint is over the normalized text — case-folded, non-
/// alphanumerics dropped, whitespace collapsed — so a rewording that changes
/// no word is the same lesson and one that changes a word is a new one (a
/// semantic near-duplicate is the reviewer's call, not this key's).
pub fn authored_dedup_key(
    family: &str,
    target_ref: &str,
    action: ActionKind,
    content: &str,
) -> String {
    format!(
        "{}\u{1f}{}",
        dedup_key(family, target_ref, action),
        content_fingerprint(content)
    )
}

/// The dedup key of a REVERT: [`dedup_key`] plus the hash of the applied
/// recommendation it retracts. An analyzer finding is "this target has this
/// kind of problem", so two findings on one target rightly collapse — but a
/// revert is about one specific applied recommendation, and two lessons on
/// the same entity that both regressed need two reverts. Until 2026-09-06
/// they shared a key and the second was dropped as a duplicate of the first
/// (`crates/areev-bench/CURVE.md`, seed 3: two regressed, one revert).
pub fn revert_dedup_key(family: &str, target_ref: &str, revert_of: &str) -> String {
    format!(
        "{}\u{1f}{}",
        dedup_key(family, target_ref, ActionKind::Revert),
        normalize_ident(revert_of)
    )
}

/// Sixteen hex chars of FNV-1a (64-bit) over the normalized content. A dedup
/// key needs stability and spread, not cryptographic strength — a collision
/// here would merge two findings in a review queue, never grant anything —
/// so this stays dependency-free, as the crate is by policy.
pub fn content_fingerprint(content: &str) -> String {
    let mut normalized = String::with_capacity(content.len());
    let mut pending_space = false;
    for c in content.chars() {
        if c.is_alphanumeric() {
            if pending_space && !normalized.is_empty() {
                normalized.push(' ');
            }
            pending_space = false;
            normalized.extend(c.to_lowercase());
        } else if c.is_whitespace() || !c.is_alphanumeric() {
            pending_space = true;
        }
    }
    let mut h: u64 = 0xcbf2_9ce4_8422_2325;
    for b in normalized.as_bytes() {
        h ^= u64::from(*b);
        h = h.wrapping_mul(0x0000_0100_0000_01b3);
    }
    format!("{h:016x}")
}

/// Lifecycle status — a rebuildable index-layer cache (the recommendation's
/// content hash is stable for its whole life).
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum RecStatus {
    #[default]
    Pending,
    Approved,
    Rejected,
    Applied,
    RolledBack,
    Expired,
}

impl RecStatus {
    pub fn as_str(&self) -> &'static str {
        match self {
            RecStatus::Pending => "pending",
            RecStatus::Approved => "approved",
            RecStatus::Rejected => "rejected",
            RecStatus::Applied => "applied",
            RecStatus::RolledBack => "rolled_back",
            RecStatus::Expired => "expired",
        }
    }

    /// Is a transition to `to` allowed from this state? `by_policy` marks the
    /// auto-apply actor, the only one permitted the reasonless
    /// `pending → applied` jump.
    pub fn can_transition_to(&self, to: RecStatus, by_policy: bool) -> bool {
        use RecStatus::*;
        match (self, to) {
            (Pending, Approved) | (Pending, Rejected) => true,
            (Pending, Applied) => by_policy, // auto-apply only
            (Approved, Applied) => true,
            (Applied, RolledBack) => true,
            // `expired` is computed from valid_to, applied to still-open recs.
            (Pending, Expired) | (Approved, Expired) => true,
            _ => false,
        }
    }
}

/// Who/what performed a transition.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum ObserverType {
    Human,
    Agent,
    Policy,
    System,
}

/// One immutable audit Observation per transition, hash-chained per
/// recommendation.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct AuditRecord {
    pub rec_hash: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub from: Option<RecStatus>,
    pub to: RecStatus,
    /// Host-asserted actor label, e.g. `user:alice`, `agent:worker-3`,
    /// `policy:auto`.
    pub actor: String,
    pub observer_type: ObserverType,
    /// Mandatory written reason (≤500 chars), the review statement's BECAUSE.
    pub because: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub previous_audit_hash: Option<String>,
    /// §7.4: present exactly when this transition applied a code revision —
    /// the evalset-run edge, recorded where the forensics look.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub gating: Option<GatingEvidence>,
    pub at_ms: i64,
}

/// Maximum length of a BECAUSE reason.
pub const MAX_BECAUSE: usize = 500;

impl AuditRecord {
    /// Build the Observation grain that records this transition. `derived_from`
    /// chains `[rec_hash, previous_audit_hash]` per the lifecycle spec.
    pub fn to_grain_spec(&self, namespace: &str) -> GrainSpec {
        let mut derived_from = vec![Value::from(self.rec_hash.clone())];
        if let Some(prev) = &self.previous_audit_hash {
            derived_from.push(Value::from(prev.clone()));
        }
        let mut spec = GrainSpec::new(crate::model::grain_type::OBSERVATION, namespace)
            .with_field("observation_kind", "loop_audit")
            .with_field("rec_hash", self.rec_hash.clone())
            .with_field("to_status", self.to.as_str())
            .with_field("actor", self.actor.clone())
            .with_field(
                "observer_type",
                serde_json::to_value(self.observer_type).unwrap(),
            )
            .with_field("because", self.because.clone())
            .with_field("at_ms", self.at_ms)
            .with_field("derived_from", Value::Array(derived_from));
        if let Some(from) = self.from {
            spec.fields
                .insert("from_status".into(), Value::from(from.as_str()));
        }
        if let Some(g) = &self.gating {
            spec.fields
                .insert("gating_evalset".into(), Value::from(g.evalset_hash.clone()));
            spec.fields
                .insert("gating_run_id".into(), Value::from(g.run_id.clone()));
            spec.fields.insert("gating_passed".into(), Value::from(g.passed));
            spec.fields.insert("gating_failed".into(), Value::from(g.failed));
        }
        spec
    }
}

/// A stored recommendation. `hash` (the content address) and `status` (the
/// index-layer cache) are set by the engine, not serialized into the grain
/// body — the body is immutable content, the lifecycle lives in the state
/// index and the audit chain.
/// One live lesson an authored lesson restates. `method` is `cosine`
/// (the substrate's embedder, T1) or `jaccard` (normalized token sets, the
/// T0 floor — weak, and honest about it).
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct NearDuplicate {
    pub hash: String,
    pub score: f64,
    pub method: String,
}

#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct Recommendation {
    #[serde(skip)]
    pub hash: String,
    pub analyzer: String,
    pub params_snapshot: Map<String, Value>,
    pub origin: Origin,
    pub target_ref: String,
    pub action_kind: ActionKind,
    pub dedup_key: String,
    pub summary: Summary,
    pub severity: Severity,
    #[serde(flatten)]
    pub proposal: Proposal,
    pub destructive: bool,
    pub rollbackable: bool,
    #[serde(default)]
    pub evidence: Vec<String>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub evidence_query: Option<String>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub metric: Option<MetricSnapshot>,
    pub confidence: f64,
    pub importance: f64,
    pub created_at_ms: i64,
    /// Optional LLM guidance: an ENRICH note on a deterministic recommendation,
    /// or an `origin = llm` draft's own rationale (§9). Whitelisted, capped
    /// text — it never replaces the engine-templated `summary`.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub guidance: Option<String>,
    /// Rule E1's pin (§7.4): the evalset hash a `code_revision` was gated
    /// against — REQUIRED for code targets, forbidden elsewhere (a
    /// recommendation targets code XOR an evalset, never both, and only
    /// code carries the pin). Shown at review; checked live at apply
    /// (superseded evalset ⇒ the pin is stale ⇒ re-gate).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub evalset_hash: Option<String>,
    /// Live lessons on the same entity this authored lesson restates in
    /// other words (`Policy::near_duplicate = flag`). The reviewer's call
    /// stays theirs — the card shows the existing rule beside the proposal.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub near_duplicate_of: Vec<NearDuplicate>,
    /// A `plan_revision`'s rehearsal against the journaled runs of the live
    /// plan (the runtime's shadow report: per-run outcome under incumbent vs
    /// candidate, effects out of support, spend) — computed at proposal
    /// time when the substrate can, so the reviewer approves from evidence,
    /// not prose. Absent when the substrate has no runtime or the plan has
    /// no journaled runs.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub replay: Option<Value>,
    #[serde(skip)]
    pub status: RecStatus,
}

/// The recorded evalset-run edge (§7.4): what an apply of a code revision
/// must present, and what its audit Observation carries. "Trust me, it
/// passed" is not an edge; (evalset, run, stats) is.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct GatingEvidence {
    /// The evalset the gate ran — must equal the recommendation's pin.
    pub evalset_hash: String,
    /// The `areev run`/`areev eval` run id that executed the gate.
    pub run_id: String,
    pub passed: u64,
    pub failed: u64,
}

/// Rule E1's structural checks (§7.4), applied when a draft is stamped and
/// re-checked when a stored grain is loaded (a hand-authored grain must not
/// bypass the rule):
/// - `code_revision` ⇔ a `tool:` target, and it MUST pin an evalset hash.
/// - `adapter_revision` ⇔ a `model:` target, same mandatory pin (the tuning
///   seam inherits §7.4's gate wholesale).
/// - An `evalset:` target is its own always-human-approved class: never a
///   gated revision, and never itself pinned (the gate cannot gate itself).
/// - No other target class carries a pin.
pub fn validate_code_rules(
    action: ActionKind,
    target_class: &str,
    evalset_hash: Option<&str>,
) -> Result<()> {
    let e1 = |why: &str| Err(Error::InvalidRecommendation(format!("Rule E1: {why}")));
    match (action, target_class) {
        (ActionKind::CodeRevision, "code") => {
            if evalset_hash.is_none_or(|h| h.trim().is_empty()) {
                return e1(
                    "a code_revision must pin the evalset hash it was gated \
                     against (evalset_hash)",
                );
            }
        }
        (ActionKind::CodeRevision, other) => {
            return e1(&format!(
                "code_revision requires a tool: target, got class '{other}'"
            ));
        }
        (ActionKind::AdapterRevision, "model") => {
            if evalset_hash.is_none_or(|h| h.trim().is_empty()) {
                return e1(
                    "an adapter_revision must pin the evalset hash it was \
                     gated against (evalset_hash)",
                );
            }
        }
        (ActionKind::AdapterRevision, other) => {
            return e1(&format!(
                "adapter_revision requires a model: target, got class '{other}'"
            ));
        }
        // A REVERT of an applied code revision targets the same tool: —
        // it is the rollback inverse, not new code, so it needs no pin
        // (blocking it here would leave the highest-stakes target class
        // unable to propose its own measured rollback).
        (ActionKind::Revert, "code") => {
            if evalset_hash.is_some() {
                return e1("a code revert carries no evalset pin");
            }
        }
        (_, "code") => {
            return e1("a tool: target requires action_kind code_revision");
        }
        // Same rollback-inverse escape hatch for the adapter class.
        (ActionKind::Revert, "model") => {
            if evalset_hash.is_some() {
                return e1("an adapter revert carries no evalset pin");
            }
        }
        (_, "model") => {
            return e1("a model: target requires action_kind adapter_revision");
        }
        (_, "evalset") => {
            if evalset_hash.is_some() {
                return e1(
                    "an evalset-target recommendation cannot pin an evalset — \
                     the gate cannot gate itself",
                );
            }
        }
        _ => {
            if evalset_hash.is_some() {
                return e1(
                    "evalset_hash is only valid on code_revision or \
                     adapter_revision recommendations",
                );
            }
        }
    }
    Ok(())
}

impl Recommendation {
    /// Serialize the immutable body into grain fields (excludes hash/status).
    pub fn to_grain_spec(&self, namespace: &str) -> Result<GrainSpec> {
        let value = serde_json::to_value(self)
            .map_err(|e| Error::Internal(format!("serialize recommendation: {e}")))?;
        let obj = value
            .as_object()
            .ok_or_else(|| Error::Internal("recommendation did not serialize to object".into()))?
            .clone();
        Ok(GrainSpec {
            grain_type: crate::model::grain_type::RECOMMENDATION.to_string(),
            namespace: namespace.to_string(),
            fields: obj,
        })
    }

    /// Reconstruct from a stored grain body. `hash` comes from the record's
    /// address; `status` is supplied from the state index by the caller.
    /// Rule E1 is RE-CHECKED here — a hand-authored recommendation grain
    /// must not smuggle an unpinned code revision past the stamped path.
    pub fn from_fields(hash: &str, fields: &Map<String, Value>) -> Result<Self> {
        let mut rec: Recommendation = serde_json::from_value(Value::Object(fields.clone()))
            .map_err(|e| Error::InvalidRecommendation(format!("decode {hash}: {e}")))?;
        rec.hash = hash.to_string();
        let class = crate::model::TargetRef::parse(&rec.target_ref)
            .map(|t| t.target_class())
            .unwrap_or("host");
        validate_code_rules(rec.action_kind, class, rec.evalset_hash.as_deref())?;
        Ok(rec)
    }
}

#[cfg(test)]
mod tests {
    /// The gate and `areev eval run --tolerance` call this one function, so
    /// the boundary is the same for both readers by construction: a drop of
    /// exactly the tolerance holds, a hair past it regresses, and a tolerance
    /// that is not a positive finite number is zero.
    #[test]
    fn the_regression_rule_holds_at_the_tolerance_and_regresses_past_it() {
        use super::is_regression;
        // Percentage points, the way `eval run --tolerance` reads it.
        assert!(!is_regression(100.0, 0.0, true, 100.0), "a 100-point drop under a 100-point tolerance is accepted");
        assert!(is_regression(100.0, 0.0, true, 99.9));
        assert!(!is_regression(92.0, 91.5, true, 1.0));
        assert!(is_regression(92.0, 90.5, true, 1.0));
        // Counts, the way the gate reads `min_effect.count`.
        assert!(!is_regression(359.0, 355.0, true, 5.0));
        assert!(!is_regression(359.0, 354.0, true, 5.0), "exactly at the floor holds");
        assert!(is_regression(359.0, 353.0, true, 5.0));
        assert!(!is_regression(28.0, 33.0, false, 5.0));
        assert!(is_regression(28.0, 34.0, false, 5.0));
        // Zero is the pre-policy rule; garbage is zero, never a wider floor.
        assert!(is_regression(359.0, 355.0, true, 0.0));
        assert!(is_regression(359.0, 355.0, true, -5.0));
        assert!(is_regression(359.0, 355.0, true, f64::NAN));
        assert!(!is_regression(1.0, 1.0, true, 0.0), "equal is never a regression");
    }

    use super::*;
    use serde_json::json;

    #[test]
    fn summary_renders_deterministically() {
        let mut args = Map::new();
        args.insert("count".into(), json!(3));
        args.insert("subject".into(), json!("acme"));
        let s = Summary::new("duplicate.exact", args);
        assert_eq!(
            s.render(),
            "Consolidate 3 exact-duplicate grains for \"acme\""
        );
    }

    #[test]
    fn dedup_key_ignores_content_and_case() {
        let a = dedup_key(
            "loop.duplicate_sweep",
            "entity:NS/John",
            ActionKind::Consolidate,
        );
        let b = dedup_key(
            "loop.duplicate_sweep",
            "entity:ns/john",
            ActionKind::Consolidate,
        );
        assert_eq!(a, b, "case-folded to one identity");
    }

    #[test]
    fn authored_dedup_key_distinguishes_content_but_not_wording_noise() {
        let a = authored_dedup_key("llm", "entity:ns/x", ActionKind::Record, "Record the vendor name.");
        let same = authored_dedup_key("llm", "entity:NS/X", ActionKind::Record, "  record THE vendor  name ");
        let other = authored_dedup_key("llm", "entity:ns/x", ActionKind::Record, "Record the amount.");
        assert_eq!(a, same, "case, punctuation and spacing are not a new lesson");
        assert_ne!(a, other, "a different lesson on the same entity is a different finding");
        assert!(a.starts_with(&dedup_key("llm", "entity:ns/x", ActionKind::Record)));
        assert_eq!(content_fingerprint("A b"), content_fingerprint("a-b"));
        assert_ne!(content_fingerprint("ab"), content_fingerprint("a b"));
    }

    #[test]
    fn a_revert_is_keyed_by_what_it_reverts() {
        let a = revert_dedup_key("loop.outcome_review", "entity:ns/capture", "aaaa");
        let b = revert_dedup_key("loop.outcome_review", "entity:ns/capture", "bbbb");
        let a2 = revert_dedup_key("loop.outcome_review", "entity:NS/Capture", "AAAA");
        assert_ne!(a, b, "two reverts on one target are two findings");
        assert_eq!(a, a2, "the same revert, case-folded, is one");
        assert!(a.starts_with(&dedup_key("loop.outcome_review", "entity:ns/capture", ActionKind::Revert)));
    }

    #[test]
    fn dedup_key_distinguishes_action() {
        let a = dedup_key("f", "entity:ns/x", ActionKind::Consolidate);
        let b = dedup_key("f", "entity:ns/x", ActionKind::FlagContradiction);
        assert_ne!(a, b);
    }

    #[test]
    fn lifecycle_gates_pending_to_applied() {
        assert!(!RecStatus::Pending.can_transition_to(RecStatus::Applied, false));
        assert!(RecStatus::Pending.can_transition_to(RecStatus::Applied, true)); // policy
        assert!(RecStatus::Pending.can_transition_to(RecStatus::Approved, false));
        assert!(RecStatus::Approved.can_transition_to(RecStatus::Applied, false));
        assert!(RecStatus::Applied.can_transition_to(RecStatus::RolledBack, false));
        assert!(!RecStatus::Rejected.can_transition_to(RecStatus::Applied, true));
    }

    #[test]
    fn rule_e1_pins_code_and_only_code() {
        use crate::model::ActionKind as A;
        // code_revision on a tool target with a pin: OK.
        assert!(validate_code_rules(A::CodeRevision, "code", Some("abc")).is_ok());
        // Unpinned code revision: refused.
        assert!(validate_code_rules(A::CodeRevision, "code", None).is_err());
        assert!(validate_code_rules(A::CodeRevision, "code", Some("  ")).is_err());
        // code_revision off a tool target: refused.
        assert!(validate_code_rules(A::CodeRevision, "memory", Some("abc")).is_err());
        // A tool target with a non-code action: refused.
        assert!(validate_code_rules(A::Flag, "code", None).is_err());
        // The gate cannot gate itself.
        assert!(validate_code_rules(A::Flag, "evalset", Some("abc")).is_err());
        assert!(validate_code_rules(A::Flag, "evalset", None).is_ok());
        // Pins don't leak onto ordinary targets.
        assert!(validate_code_rules(A::Consolidate, "memory", Some("abc")).is_err());
        assert!(validate_code_rules(A::Consolidate, "memory", None).is_ok());
    }

    #[test]
    fn rule_e1_pins_adapter_revisions_like_code() {
        use crate::model::ActionKind as A;
        // adapter_revision on a model target with a pin: OK.
        assert!(validate_code_rules(A::AdapterRevision, "model", Some("abc")).is_ok());
        // Unpinned adapter revision: refused.
        assert!(validate_code_rules(A::AdapterRevision, "model", None).is_err());
        assert!(validate_code_rules(A::AdapterRevision, "model", Some("  ")).is_err());
        // adapter_revision off a model target: refused.
        assert!(validate_code_rules(A::AdapterRevision, "memory", Some("abc")).is_err());
        assert!(validate_code_rules(A::AdapterRevision, "code", Some("abc")).is_err());
        // A model target with a non-adapter action: refused.
        assert!(validate_code_rules(A::Flag, "model", None).is_err());
        // The rollback inverse needs no pin — and must carry none.
        assert!(validate_code_rules(A::Revert, "model", None).is_ok());
        assert!(validate_code_rules(A::Revert, "model", Some("abc")).is_err());
    }

    #[test]
    fn from_fields_rechecks_rule_e1() {
        // A hand-authored grain body carrying an unpinned code revision must
        // not decode into a usable recommendation.
        let mut rec = Recommendation {
            hash: String::new(),
            analyzer: "loop.codegen/1".into(),
            params_snapshot: Map::new(),
            origin: Origin::Builtin,
            target_ref: "tool:abc123".into(),
            action_kind: ActionKind::CodeRevision,
            dedup_key: "k".into(),
            summary: Summary::new("command.finding", Map::new()),
            severity: Severity::Low,
            proposal: Proposal::Data { data: Map::new() },
            destructive: false,
            rollbackable: false,
            evidence: vec![],
            evidence_query: None,
            metric: None,
            confidence: 0.5,
            importance: 0.5,
            created_at_ms: 0,
            guidance: None,
            evalset_hash: None, // the smuggle attempt
            near_duplicate_of: Vec::new(),
            replay: None,
            status: RecStatus::Pending,
        };
        let spec = rec.to_grain_spec("ns").unwrap();
        assert!(Recommendation::from_fields("h", &spec.fields).is_err());
        rec.evalset_hash = Some("es-hash".into());
        let spec = rec.to_grain_spec("ns").unwrap();
        let back = Recommendation::from_fields("h", &spec.fields).unwrap();
        assert_eq!(back.evalset_hash.as_deref(), Some("es-hash"));
    }

    #[test]
    fn recommendation_round_trips_through_fields() {
        let rec = Recommendation {
            hash: "ignored".into(),
            analyzer: "loop.staleness/1".into(),
            params_snapshot: Map::new(),
            origin: Origin::Builtin,
            target_ref: "grain:sha256:abc".into(),
            action_kind: ActionKind::Expire,
            dedup_key: "k".into(),
            summary: Summary::new("staleness.expired", Map::new()),
            severity: Severity::Low,
            proposal: Proposal::Cal {
                cal: "FORGET sha256:abc".into(),
            },
            destructive: true,
            rollbackable: false,
            evidence: vec!["sha256:abc".into()],
            evidence_query: None,
            metric: None,
            confidence: 0.9,
            importance: 0.4,
            created_at_ms: 1000,
            guidance: None,
            evalset_hash: None,
            near_duplicate_of: Vec::new(),
            replay: None,
            status: RecStatus::Pending,
        };
        let spec = rec.to_grain_spec("ns").unwrap();
        // hash and status are excluded from the immutable body.
        assert!(!spec.fields.contains_key("hash"));
        assert!(!spec.fields.contains_key("status"));
        let back = Recommendation::from_fields("realhash", &spec.fields).unwrap();
        assert_eq!(back.hash, "realhash");
        assert_eq!(back.analyzer, "loop.staleness/1");
        assert!(back.destructive);
        assert!(matches!(back.proposal, Proposal::Cal { .. }));
    }
}