package signature

import (
	"context"

	"github.com/terst/listen/server/internal/advertising"
)

// Audio supplies decoded audio at [SampleRate].
type Audio interface {
	// Episode decodes a whole episode.
	Episode(ctx context.Context, id string) ([]float64, error)
	// Range decodes one stretch of an episode.
	Range(ctx context.Context, id string, startMs, durationMs int64) ([]float64, error)
}

// ShadowBreak is one stretch detection decided about.
type ShadowBreak struct {
	StartMs int64
	EndMs   int64
	// Partners is how many other episodes detection found carrying this audio. It is what
	// makes an advertisement a confirmed repeat rather than a single sighting.
	Partners int
	// Kind is what skips.Classify called it: "ad", or one of the show's own "intro", "cue",
	// "outro".
	Kind string
}

// ShadowLabel is something the listener said about a stretch of this episode.
type ShadowLabel struct {
	StartMs int64
	EndMs   int64
	// Advertising is true when he said this is an ad -- the notification button, or marking a
	// break the detector found as right. False when he said it is the programme.
	Advertising bool
	// EndMeasured is false when the end is wherever the jump landed rather than where he heard
	// the advertising stop. See [GuessedEndSeedMs].
	EndMeasured bool
}

// ShadowEpisode is one episode of a batch, what detection made of it, and what the listener
// said about it.
type ShadowEpisode struct {
	ID      string
	FeedURL string
	Breaks  []ShadowBreak
	Labels  []ShadowLabel
}

// ShadowBatch is one detection pass over one show.
type ShadowBatch struct {
	// Occasion identifies the pass. Everything in it counts as a single sighting, because the
	// episodes were mirrored together and all carry whatever campaign was running that minute.
	Occasion string
	AtUnix   int64
	Episodes []ShadowEpisode
}

// Proposal is a stretch the library would cut.
type Proposal struct {
	EpisodeID string
	StartMs   int64
	EndMs     int64
	Entry     *Entry
	Votes     int
	// Novel is true when detection did not already publish a break covering this. Those are
	// the interesting ones: the library earning its place rather than agreeing with what is
	// already known.
	Novel bool
}

// GuessedEndSeedMs is how much of a reported advertisement is kept when its end was never
// measured.
//
// An uncorrected end is wherever the 25-second jump happened to land, not where the advertising
// stopped, so the tail of that range may be programme -- and a signature carrying programme is
// the one thing this must never build. One advertising slot from the start is far more audio
// than a landmark fingerprint needs to be distinctive, and it is the part most certainly
// advertising.
const GuessedEndSeedMs = advertising.UnitMs

// Report is what one shadow run saw.
type Report struct {
	Harvested  int
	Reinforced int
	// Disowned counts signatures dropped because the listener said that audio is the programme.
	Disowned int
	WouldCut []Proposal
	// Skipped counts episodes whose audio could not be read. Reported rather than failing the
	// run: one unreadable file must not cost the whole batch's learning.
	Skipped int
}

// Shadow learns from a detection pass and reports what the library would do, without doing it.
//
// Two jobs, deliberately in one place. It harvests signatures from what detection found, and it
// matches what it already holds against the same audio. Running those together is what makes the
// log honest: the proposals in a report were made by a library that did not yet know about this
// batch.
type Shadow struct {
	library *Library
	audio   Audio
	// MinSightings is how many separate occasions a signature needs before it is matched
	// against. Never one: a batch is mirrored in a minute and shares one campaign.
	MinSightings int
}

// NewShadow returns a shadow runner over the given library.
func NewShadow(library *Library, audio Audio) *Shadow {
	return &Shadow{library: library, audio: audio, MinSightings: 2}
}

// Run matches first, then harvests.
//
// The order is the point. Harvesting first would let this batch's own signatures match the batch
// they came from, and every run would report itself as a triumph.
func (s *Shadow) Run(ctx context.Context, batch ShadowBatch) (Report, error) {
	var report Report
	// Counted per episode, not per failure: an episode that could not be read fails both the
	// matching pass and the harvesting pass, and reporting it twice would overstate the damage.
	unreadable := map[string]bool{}
	for _, ep := range batch.Episodes {
		if ctx.Err() != nil {
			report.Skipped = len(unreadable)
			return report, ctx.Err()
		}
		samples, err := s.audio.Episode(ctx, ep.ID)
		if err != nil || len(samples) == 0 {
			unreadable[ep.ID] = true
			continue
		}
		for _, m := range s.library.Match(Take(samples), s.MinSightings) {
			report.WouldCut = append(report.WouldCut, Proposal{
				EpisodeID: ep.ID,
				StartMs:   m.StartMs,
				EndMs:     m.EndMs,
				Entry:     m.Entry,
				Votes:     m.Votes,
				Novel:     !covered(ep.Breaks, m.StartMs, m.EndMs),
			})
		}
	}

	for _, ep := range batch.Episodes {
		if ctx.Err() != nil {
			report.Skipped = len(unreadable)
			return report, ctx.Err()
		}
		// His "no" first, and it does two things: it stops the harvest below from storing
		// audio he says is the programme, and it takes back anything already held that matches
		// it. A signature built from programme would cut the programme in every episode
		// carrying it, until somebody noticed.
		for _, label := range ep.Labels {
			if label.Advertising {
				continue
			}
			if !s.disown(ctx, ep, label, &report) {
				unreadable[ep.ID] = true
			}
		}
		for _, br := range ep.Breaks {
			if disowned(ep.Labels, br.StartMs, br.EndMs) {
				continue
			}
			// A break this cannot read is counted and stepped over. Every other per-item loop
			// in this server isolates a single failure the same way, and the first real run
			// lost a whole show's learning to one episode whose path was wrong.
			if !s.harvest(ctx, batch, ep, seed{
				startMs:   br.StartMs,
				endMs:     br.EndMs,
				kind:      kindFor(br.Kind),
				partners:  br.Partners,
				confirmed: vouchedFor(ep.Labels, br.StartMs, br.EndMs),
			}, &report) {
				unreadable[ep.ID] = true
				break
			}
		}
		// Then what he pointed at himself. Advertising detection missed is the gap the library
		// exists to close, so a report is harvested whether or not a break was found there.
		for _, label := range ep.Labels {
			if !label.Advertising || coveredByBreak(ep.Breaks, label) {
				continue
			}
			if !s.harvest(ctx, batch, ep, seedFromLabel(label), &report) {
				unreadable[ep.ID] = true
				break
			}
		}
	}
	report.Skipped = len(unreadable)
	return report, nil
}

// seed is one stretch about to become a signature.
type seed struct {
	startMs   int64
	endMs     int64
	kind      string
	partners  int
	confirmed bool
}

func seedFromLabel(label ShadowLabel) seed {
	end := label.EndMs
	if !label.EndMeasured && end-label.StartMs > GuessedEndSeedMs {
		end = label.StartMs + GuessedEndSeedMs
	}
	// Always a copy: he presses the button when he hears an advertisement, and a show's own
	// furniture is not what that button is for.
	return seed{startMs: label.StartMs, endMs: end, kind: KindCopy, confirmed: true}
}

func overlaps(aStart, aEnd, bStart, bEnd int64) bool { return aStart < bEnd && aEnd > bStart }

// disowned says whether the listener called this stretch the programme.
func disowned(labels []ShadowLabel, startMs, endMs int64) bool {
	for _, l := range labels {
		if !l.Advertising && overlaps(startMs, endMs, l.StartMs, l.EndMs) {
			return true
		}
	}
	return false
}

// vouchedFor says whether the listener marked this break right.
func vouchedFor(labels []ShadowLabel, startMs, endMs int64) bool {
	for _, l := range labels {
		if l.Advertising && overlaps(startMs, endMs, l.StartMs, l.EndMs) {
			return true
		}
	}
	return false
}

// coveredByBreak says whether detection already found this label's audio, in which case
// harvesting the break covered it and the label only had to vouch for it.
func coveredByBreak(breaks []ShadowBreak, label ShadowLabel) bool {
	for _, br := range breaks {
		if overlaps(label.StartMs, label.EndMs, br.StartMs, br.EndMs) {
			return true
		}
	}
	return false
}

// disown forgets every signature matching audio the listener says is the programme.
func (s *Shadow) disown(ctx context.Context, ep ShadowEpisode, label ShadowLabel, report *Report) bool {
	samples, err := s.audio.Range(ctx, ep.ID, label.StartMs, label.EndMs-label.StartMs)
	if err != nil {
		return false
	}
	report.Disowned += s.library.Disown(Take(samples))
	return true
}

// harvest stores one stretch of audio as a signature, and reports whether the episode could be
// read at all. A false means give up on this episode, not on the batch.
func (s *Shadow) harvest(ctx context.Context, batch ShadowBatch, ep ShadowEpisode, sd seed, report *Report) bool {
	duration := sd.endMs - sd.startMs
	if duration <= 0 {
		return true
	}
	samples, err := s.audio.Range(ctx, ep.ID, sd.startMs, duration)
	if err != nil {
		return false
	}
	fp := Take(samples)
	if len(fp.Hashes) == 0 {
		return true
	}
	_, isNew := s.library.Observe(Observation{
		Fingerprint: fp,
		Kind:        sd.kind,
		FeedURL:     ep.FeedURL,
		DurationMs:  duration,
		Partners:    sd.partners,
		Confirmed:   sd.confirmed,
		Occasion:    batch.Occasion,
	}, batch.AtUnix)
	if isNew {
		report.Harvested++
	} else {
		report.Reinforced++
	}
	return true
}

// kindFor maps what detection called a break onto what the library stores.
//
// Advertising belongs to nobody and expires; everything else is the show's own audio and does
// not. Anything unrecognised is treated as advertising, matching the app's rule that a kind a
// newer server invented must not quietly become permanent furniture.
func kindFor(breakKind string) string {
	switch breakKind {
	case "intro", "cue", "outro":
		return KindCue
	default:
		return KindCopy
	}
}

// covered says whether detection already published a break overlapping this stretch.
func covered(breaks []ShadowBreak, startMs, endMs int64) bool {
	for _, br := range breaks {
		if startMs < br.EndMs && endMs > br.StartMs {
			return true
		}
	}
	return false
}