Maven/internal/router/eval/eval.go

// Package eval is the held-out routing contract — the fixture Vikunja #319
// measures against before #320 flips the default route decider.
//
// Why it is a separate package from router: the fixture must be scorable by
// BOTH paths (today's classifier cascade and the resident model's LLM router)
// from outside the router package, including from cmd/mavend where the real
// llama-server client lives. A _test.go file in router can't be imported, and
// testdata isn't reachable from another package's working directory — so the
// fixture is embedded here and the scorer takes a Router interface.
//
// The fixture is HELD OUT from models/seeds/*.txt on purpose: a classifier
// scored on its own seed phrases measures memorisation of frozen centroids,
// which is exactly the illusion that hid the weak RU query handling. See
// TestFixtureIsHeldOut, which enforces it.
package eval

import (
	"context"
	_ "embed"
	"encoding/json"
	"fmt"
	"sort"
	"strings"
	"time"

	"github.com/kami/maven/internal/router"
)

//go:embed ru_routing_v1.json
var fixtureJSON []byte

// Case — one utterance and the route it must produce. Slot expectations are
// deliberately coarse (see the fixture's notes): want_fn is a boolean because
// the fn allowlist lives in deploy config, and want_fact_key names the loop's
// rule keys because a fact under the wrong key starves its predicate.
//
// Intent is empty exactly when WantClarify is set: the contract there is that
// the router refuses instead of guessing.
type Case struct {
	ID          string        `json:"id"`
	Utterance   string        `json:"utterance"`
	Lang        string        `json:"lang"`
	Intent      router.Intent `json:"intent"`
	WantTime    bool          `json:"want_time"`
	WantFn      bool          `json:"want_fn"`
	WantFactKey string        `json:"want_fact_key"`
	WantClarify bool          `json:"want_clarify"`
	Tags        []string      `json:"tags"`
	Note        string        `json:"note"`
}

// Fixture — the versioned envelope, same shape as
// cmd/mavend/testdata/system_safety_scenarios.json. SchemaVersion gates the
// loader so an older binary refuses a fixture it would misread rather than
// scoring it wrong and reporting a number.
type Fixture struct {
	SchemaVersion int      `json:"schema_version"`
	Name          string   `json:"name"`
	ReferenceNow  string   `json:"reference_now"`
	Notes         []string `json:"notes"`
	Cases         []Case   `json:"cases"`
}

// SchemaVersion — the version this package understands.
const SchemaVersion = 1

// Load returns the embedded fixture.
func Load() (Fixture, error) {
	var f Fixture
	if err := json.Unmarshal(fixtureJSON, &f); err != nil {
		return Fixture{}, fmt.Errorf("parse fixture: %w", err)
	}
	if f.SchemaVersion != SchemaVersion {
		return Fixture{}, fmt.Errorf("fixture schema_version %d, want %d", f.SchemaVersion, SchemaVersion)
	}
	if len(f.Cases) == 0 {
		return Fixture{}, fmt.Errorf("fixture has no cases")
	}
	return f, nil
}

// Now — the fixture's reference clock. Relative reminder slots ("через
// полчаса") resolve against it, so a scoring run is reproducible regardless of
// when it runs.
func (f Fixture) Now() (time.Time, error) {
	t, err := time.Parse(time.RFC3339, f.ReferenceNow)
	if err != nil {
		return time.Time{}, fmt.Errorf("parse reference_now %q: %w", f.ReferenceNow, err)
	}
	return t, nil
}

// Router — the one thing a route decider must do to be scorable. *router.Router
// satisfies it directly; an LLM-only path wraps its Route in RouterFunc.
type Router interface {
	Route(ctx context.Context, utterance string, now time.Time) (router.Decision, error)
}

// RouterFunc adapts a bare function to Router — for scoring a single stage
// (e.g. *router.LLMRouter, whose Route returns an extra ok bool) without
// standing up the whole cascade.
type RouterFunc func(ctx context.Context, utterance string, now time.Time) (router.Decision, error)

// Route implements Router.
func (f RouterFunc) Route(ctx context.Context, utterance string, now time.Time) (router.Decision, error) {
	return f(ctx, utterance, now)
}

// Outcome — one scored case. Reasons is empty exactly when Pass is true.
type Outcome struct {
	Case     Case
	Decision router.Decision
	Err      error
	Latency  time.Duration
	Pass     bool
	// IntentOK is tracked separately from Pass: a case can land the right
	// intent and still fail on a slot, and #319 needs those two numbers apart
	// (a slot gap is a parser fix; a wrong intent is a router fix).
	IntentOK bool
	Reasons  []string
}

// Report — the aggregate. Accuracy is the headline; the rest exists so a
// regression names itself instead of just moving a percentage.
type Report struct {
	Name      string
	Total     int
	Passed    int
	IntentHit int
	// FalseClarify — the router asked when the fixture expected a decision.
	// A gap, recoverable by asking again.
	FalseClarify int
	// MissedClarify — the router decided confidently where the fixture
	// expected a refusal. The dangerous direction: "сделай это" routed to an
	// act is a confident destructive guess.
	MissedClarify int
	Errors        int
	// SlotsDeferred — stage-0 hits whose slot the daemon fills downstream
	// (reminder grammar → applyAction's time parser). Not a miss, but not a
	// full router-level win either; tracked so the two aren't conflated.
	SlotsDeferred int
	Outcomes      []Outcome
	// Confusion counts want→got intent pairs, decided cases only.
	Confusion map[string]int
	// ByTag accuracy for the fixture's tags ("hard", "homelab", …).
	ByTag map[string]TagStat
	// ByLang accuracy — the RU/EN split is the whole reason this fixture
	// exists.
	ByLang map[string]TagStat
	P50    time.Duration
	P95    time.Duration
	Max    time.Duration
}

// TagStat — passed/total for one slice of the fixture.
type TagStat struct{ Passed, Total int }

// Accuracy — fraction of cases fully satisfied (intent AND slots AND the
// clarify contract).
func (r Report) Accuracy() float64 {
	if r.Total == 0 {
		return 0
	}
	return float64(r.Passed) / float64(r.Total)
}

// IntentAccuracy — fraction with the right intent, ignoring slot fills.
func (r Report) IntentAccuracy() float64 {
	if r.Total == 0 {
		return 0
	}
	return float64(r.IntentHit) / float64(r.Total)
}

// Score runs every case through r and aggregates. It never fails the run on a
// route error — an erroring case scores as a miss and is counted in Errors,
// because "the model was down" and "the model was wrong" are different numbers
// and #319 needs to tell them apart.
//
// Latency is wall-clock per Route call, including any llama-server round trip.
// That is the point on the CPU-only target: a correctness win the resident
// model pays for with seconds per turn is not a win.
func Score(ctx context.Context, name string, r Router, f Fixture) (Report, error) {
	now, err := f.Now()
	if err != nil {
		return Report{}, err
	}
	rep := Report{
		Name:      name,
		Total:     len(f.Cases),
		Confusion: map[string]int{},
		ByTag:     map[string]TagStat{},
		ByLang:    map[string]TagStat{},
	}
	lat := make([]time.Duration, 0, len(f.Cases))

	for _, c := range f.Cases {
		start := time.Now()
		d, err := r.Route(ctx, c.Utterance, now)
		o := Outcome{Case: c, Decision: d, Err: err, Latency: time.Since(start)}
		lat = append(lat, o.Latency)

		switch {
		case err != nil:
			rep.Errors++
			o.Reasons = append(o.Reasons, fmt.Sprintf("route error: %v", err))
		case c.WantClarify:
			// Only the refusal matters here; whatever intent the cascade
			// guessed underneath is irrelevant if it gated.
			o.IntentOK = d.Clarify
			if !d.Clarify {
				rep.MissedClarify++
				o.Reasons = append(o.Reasons, fmt.Sprintf("decided %q confidently (%.3f), want clarify", d.Intent, d.Confidence))
			}
		default:
			o.IntentOK = d.Intent == c.Intent && !d.Clarify
			if d.Clarify {
				rep.FalseClarify++
				o.Reasons = append(o.Reasons, fmt.Sprintf("clarified (%.3f), want intent %q", d.Confidence, c.Intent))
			} else if d.Intent != c.Intent {
				rep.Confusion[string(c.Intent)+"→"+string(d.Intent)]++
				o.Reasons = append(o.Reasons, fmt.Sprintf("intent %q, want %q (%.3f)", d.Intent, c.Intent, d.Confidence))
			}
			if c.WantTime && !d.Slots.HasTime {
				// Stage 0 skips the extractor by design: ReminderGrammar
				// captures the text after "напомни"/"remind me" and the
				// daemon's applyAction runs the time parser on it (see
				// stage0.go). Charging the router for a slot it was never
				// asked to fill would make an exact-match win look like a
				// miss — so it is counted, not failed.
				if d.Stage == 0 {
					rep.SlotsDeferred++
				} else {
					o.Reasons = append(o.Reasons, "no time slot, want one")
				}
			}
			if c.WantFn && !d.Slots.HasFn {
				o.Reasons = append(o.Reasons, "no fn slot, want an allowlist match")
			}
			if c.WantFactKey != "" && d.Slots.Key != c.WantFactKey {
				o.Reasons = append(o.Reasons, fmt.Sprintf("fact key %q, want %q", d.Slots.Key, c.WantFactKey))
			}
		}

		o.Pass = len(o.Reasons) == 0
		if o.Pass {
			rep.Passed++
		}
		if o.IntentOK {
			rep.IntentHit++
		}
		bump(rep.ByLang, c.Lang, o.Pass)
		for _, tag := range c.Tags {
			bump(rep.ByTag, tag, o.Pass)
		}
		rep.Outcomes = append(rep.Outcomes, o)
	}

	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
	rep.P50, rep.P95 = percentile(lat, 0.50), percentile(lat, 0.95)
	if len(lat) > 0 {
		rep.Max = lat[len(lat)-1]
	}
	return rep, nil
}

func bump(m map[string]TagStat, key string, pass bool) {
	if key == "" {
		return
	}
	s := m[key]
	s.Total++
	if pass {
		s.Passed++
	}
	m[key] = s
}

// percentile — nearest-rank on a pre-sorted slice. No interpolation: with ~80
// samples an interpolated p95 invents a latency no turn actually took.
func percentile(sorted []time.Duration, p float64) time.Duration {
	if len(sorted) == 0 {
		return 0
	}
	i := int(p * float64(len(sorted)))
	if i >= len(sorted) {
		i = len(sorted) - 1
	}
	return sorted[i]
}

// String renders the report as the comparison table #319 pastes into the task —
// headline accuracy, the two clarify directions apart, latency, and the slices
// that name where a path is weak.
func (r Report) String() string {
	var b strings.Builder
	fmt.Fprintf(&b, "%s: %d/%d cases (%.1f%% full, %.1f%% intent-only)\n",
		r.Name, r.Passed, r.Total, 100*r.Accuracy(), 100*r.IntentAccuracy())
	fmt.Fprintf(&b, "  clarify: %d false (asked, shouldn't) / %d missed (guessed, shouldn't) | errors: %d | slots deferred to daemon: %d\n",
		r.FalseClarify, r.MissedClarify, r.Errors, r.SlotsDeferred)
	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
	fmt.Fprintf(&b, "  by lang: %s\n", renderStats(r.ByLang))
	fmt.Fprintf(&b, "  by tag:  %s\n", renderStats(r.ByTag))
	if len(r.Confusion) > 0 {
		fmt.Fprintf(&b, "  confusion: %s\n", renderCounts(r.Confusion))
	}
	return b.String()
}

// Failures — the per-case detail, sorted by ID so two runs diff cleanly.
func (r Report) Failures() string {
	var b strings.Builder
	out := append([]Outcome(nil), r.Outcomes...)
	sort.Slice(out, func(i, j int) bool { return out[i].Case.ID < out[j].Case.ID })
	for _, o := range out {
		if o.Pass {
			continue
		}
		fmt.Fprintf(&b, "  %s %q: %s\n", o.Case.ID, o.Case.Utterance, strings.Join(o.Reasons, "; "))
	}
	return b.String()
}

func renderStats(m map[string]TagStat) string {
	keys := make([]string, 0, len(m))
	for k := range m {
		keys = append(keys, k)
	}
	sort.Strings(keys)
	parts := make([]string, 0, len(keys))
	for _, k := range keys {
		s := m[k]
		parts = append(parts, fmt.Sprintf("%s %d/%d", k, s.Passed, s.Total))
	}
	return strings.Join(parts, "  ")
}

func renderCounts(m map[string]int) string {
	keys := make([]string, 0, len(m))
	for k := range m {
		keys = append(keys, k)
	}
	sort.Slice(keys, func(i, j int) bool {
		if m[keys[i]] != m[keys[j]] {
			return m[keys[i]] > m[keys[j]]
		}
		return keys[i] < keys[j]
	})
	parts := make([]string, 0, len(keys))
	for _, k := range keys {
		parts = append(parts, fmt.Sprintf("%s ×%d", k, m[k]))
	}
	return strings.Join(parts, "  ")
}