textmachine/backend/internal/pipeline/offtarget_corpus_test.go

package pipeline

import (
	"crypto/sha256"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"os"
	"sort"
	"strings"
	"testing"

	"textmachine/backend/internal/lang"
	"textmachine/backend/internal/langscreen"
	"textmachine/backend/internal/llm"
)

// offtarget_corpus_test.go: backlog row 113 — the numbers the whole target-language screen (row 46) is
// argued from were produced by a TEMPORARY harness in this package that the session deleted, so «18 of 20
// caught, 0 false positives on 1055, 13× margin» went into a contract with nothing left that could
// reproduce it. This is the base any proposal for row 46 has to beat, and it lives in the repository now.
//
// HOW THE CORPUS WAS TAKEN. Every checkpoint in every stand database under `<repo>/books/gu-zhenren`
// (162 files), put through the engine's OWN preprocessing — splitBanknote then checks.StripThink, which is
// the text the gate actually sees (stagerun cuts the banknote, then classifies) — deduplicated by the
// hash of that text, and restricted to the SHIPPING roles: draft/translator and edit/editor. Bank-role
// replies are excluded because their answers legitimately consist of source terms and the runner exempts
// them (SourceEchoExpected); including them would measure the exemption, not the gate. That leaves 1162
// outputs, of which 30 are off-target and 1132 are healthy.
//
// HOW A SPECIMEN WAS LABELLED, and why not by the gate. An output is off-target when its share of
// CYRILLIC letters is below 0.5 — a property of the TARGET side, independent of the source-script gate
// under test, so the label cannot be a restatement of the verdict. The corpus separates completely on it:
// every healthy output is above 0.9 and every off-target one is at 0.000, with the band between them
// EMPTY. Two of the thirty are fluent English (4690 and 4811 letters — the same two the original
// measurement names), one is Japanese, twenty-seven are Chinese echoes.
//
// ⚠ THE FIXTURE IS A SAMPLE OF THAT CORPUS, NOT THE CORPUS, and the numbers below are established over the
// sample. All THIRTY off-target specimens are kept — the class is scarce and every member is a distinct
// failure — while the healthy class is drawn deliberately hard rather than at random: the twenty with the
// HIGHEST source-script share (nearest the echo gate), the fifteen with the LOWEST target-script share
// (nearest the target screen), the three that fall under the evidence floor, and a spread of the rest. So
// «0 false positives» here means «0 on the forty-four hardest of the 1132», which is the useful claim; it
// is a weaker statement than 0 on all 1132, and saying otherwise would be the over-claim this header exists
// to prevent.
//
// ⚠ WHAT THE FIXTURE CARRIES, AND WHY NOT PROSE. Each row is the specimen's rune ACCOUNTING —
// total runes, runes in the source script, letters — plus a sixty-rune head for a human, its provenance,
// and its label. It is deliberately not the texts: those are a copyrighted web novel and its machine
// translations, which this repository keeps out of git on purpose (the stand lives in a separate
// repository; `example/` ships an own-authored fragment for the same reason). The accounting is
// sufficient BY CONSTRUCTION for what is under test — sourceScriptShare is a ratio of rune classes and
// reads nothing else — and the tests below feed the real function a script-faithful surrogate built from
// the counts, so the function stays exercised rather than replaced by arithmetic.
// The named LOSS: a future detector that reads words, n-grams or orthography cannot be measured on this
// fixture, and re-taking the corpus (the walk above) is what that would need.

const offTargetCorpusFile = "testdata/offtarget-corpus.json"

type offTargetSpecimen struct {
	ID                string `json:"id"`
	Label             string `json:"label"` // off_target | healthy
	SourceScriptRunes int    `json:"source_script_runes"`
	TotalRunes        int    `json:"total_runes"`
	Letters           int    `json:"letters"`
	// TargetScriptLetters is the other half of the accounting: how many of Letters are written in the
	// TARGET's script. It is what the row-46 screen weighs, and it is recorded beside the source-side
	// counts so one corpus answers both questions and the two can never drift onto different samples.
	TargetScriptLetters int    `json:"target_script_letters"`
	Provenance          string `json:"provenance"`
	Head                string `json:"head"`
}

// share is the specimen's recorded source-script share.
func (s offTargetSpecimen) share() float64 {
	return float64(s.SourceScriptRunes) / float64(s.TotalRunes)
}

// surrogate renders the specimen's rune accounting as a string the real detector can read: exactly
// SourceScriptRunes Han runes followed by exactly (TotalRunes-SourceScriptRunes) Cyrillic ones. Any two
// runes of the right classes would do — the function counts classes, not characters.
func (s offTargetSpecimen) surrogate() string {
	var b strings.Builder
	b.Grow(s.TotalRunes * 3)
	for i := 0; i < s.SourceScriptRunes; i++ {
		b.WriteRune('源')
	}
	for i := s.SourceScriptRunes; i < s.TotalRunes; i++ {
		b.WriteRune('а')
	}
	return b.String()
}

// targetShare is the specimen's recorded target-script share, over letters.
func (s offTargetSpecimen) targetShare() float64 {
	if s.Letters == 0 {
		return 0
	}
	return float64(s.TargetScriptLetters) / float64(s.Letters)
}

// targetSurrogate renders the specimen's LETTER accounting for the target screen the way surrogate() does
// for the echo gate: exactly TargetScriptLetters Cyrillic letters and the rest Latin, so the real screen
// weighs the specimen's real proportions.
func (s offTargetSpecimen) targetSurrogate() string {
	var b strings.Builder
	b.Grow(s.Letters * 2)
	for i := 0; i < s.TargetScriptLetters; i++ {
		b.WriteRune('а')
	}
	for i := s.TargetScriptLetters; i < s.Letters; i++ {
		b.WriteRune('a')
	}
	return b.String()
}

func loadOffTargetCorpus(t *testing.T) []offTargetSpecimen {
	t.Helper()
	raw, err := os.ReadFile(offTargetCorpusFile)
	if err != nil {
		t.Fatalf("read the off-target corpus fixture: %v", err)
	}
	var out []offTargetSpecimen
	if err := json.Unmarshal(raw, &out); err != nil {
		t.Fatalf("the off-target corpus fixture is not readable JSON: %v", err)
	}
	if len(out) == 0 {
		t.Fatal("the fixture is empty — this test would be enforcing nothing")
	}
	return out
}

// TestTheOffTargetCorpusIsWhatItClaims guards the fixture itself before anything is measured on it. A
// corpus that quietly lost its off-target class, or whose labels stopped separating, would make every
// number below look excellent.
func TestTheOffTargetCorpusIsWhatItClaims(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	counts := map[string]int{}
	seen := map[string]bool{}
	for _, s := range corpus {
		counts[s.Label]++
		if seen[s.ID] {
			t.Errorf("duplicate specimen %s — a repeated row would weight one output twice", s.ID)
		}
		seen[s.ID] = true
		if s.TotalRunes <= 0 || s.SourceScriptRunes < 0 || s.SourceScriptRunes > s.TotalRunes {
			t.Errorf("%s: impossible accounting %+v", s.ID, s)
		}
		if s.Letters > s.TotalRunes {
			t.Errorf("%s: more letters than runes %+v", s.ID, s)
		}
		if s.TargetScriptLetters < 0 || s.TargetScriptLetters > s.Letters {
			t.Errorf("%s: impossible target accounting %+v", s.ID, s)
		}
		if s.Provenance == "" {
			t.Errorf("%s: a specimen with no provenance cannot be re-taken", s.ID)
		}
	}
	if counts["off_target"] < 20 {
		t.Errorf("the off-target class is the scarce one and the fixture must keep it: %d rows", counts["off_target"])
	}
	if counts["healthy"] < 40 {
		t.Errorf("the healthy class is the false-positive denominator: %d rows", counts["healthy"])
	}
	if len(counts) != 2 {
		t.Errorf("labels must be exactly {off_target, healthy}, got %v", counts)
	}
}

// TestTheCorpusStillHoldsItsHardestSpecimens is the fixture's LEDGER, and it exists because a floor is not
// a measurement. The margin test asserts «at least 3×», which is the STOP row 113 names — and a fixture
// quietly softened until the hard cases were gone would sail through it reporting a BETTER number than the
// corpus ever had. Found by planting into the DATA: zeroing the source runes of the single worst healthy
// specimen left the whole battery green while making the margin larger.
//
// So the numbers the contract is argued from are RECORDED here, exactly as the prompt-label ledger records
// a prompt's bytes, and for the same reason: what must never happen silently is the corpus and the claim
// drifting apart. ⚠ RE-TAKING THE CORPUS IS LEGITIMATE, and these numbers are then re-recorded in the same
// commit — the point is that it costs an edit here and a sentence in the report, not nothing.
func TestTheCorpusStillHoldsItsHardestSpecimens(t *testing.T) {
	// Measured over 1162 shipping completions of the stand (162 databases), of which the fixture keeps all
	// 30 off-target and the 44 hardest healthy. See the file header for how the healthy class was drawn.
	const (
		wantOffTargets       = 30
		wantHealthy          = 44
		wantWorstHealthySrc  = 0.011655 // f3367ee44f56 — the specimen the 12.87× margin rests on
		wantWorstHealthyTgt  = 0.976706 // 905834e2606c — the specimen the 0.50 floor is placed under
		wantShortestOffChars = 346      // 1612c74012cc — what the 200-letter evidence floor is calibrated against
		tol                  = 1e-6
		// wantBestOffTargetTgt is the OTHER side of the interval, and its absence was a hole: the screen's
		// doc comment argues «no near-miss at all on the other side» and the threshold test compares that
		// claim's constant against another constant. The data itself was unpinned — planted, an English
		// draft's target share moved 0.000 → 0.490 and every corpus test stayed green, which would have left
		// the 0.50 floor sitting directly on top of a specimen while the ledger said the band was empty.
		wantBestOffTargetTgt = 0.0
		// corpusDigest is the POPULATION, not its extrema. The ledger above records one number per axis, so
		// every specimen except the worst one could be softened silently: planted, the second-worst healthy
		// source share went 42 → 0 runes, and separately a hard row was overwritten with an easy row's
		// accounting under its own id — both invisible, because the extremum still held. The corpus is a
		// frozen measurement, so a digest over every row's accounting is the honest pin: re-taking it is a
		// deliberate act that re-records this line in the same commit.
		// ⚠ IT COVERS EVERY FIELD, INCLUDING THE ONES NO MEASUREMENT READS. The first version folded only the
		// six counting fields, and a planting that rewrote a specimen's `head` — and, worse, its
		// `provenance` — passed with the whole ledger green. `provenance` is the ADDRESS a specimen was
		// taken from and the address it would be re-taken from; a corpus whose numbers are right and whose
		// provenance is wrong is unreproducible in the one way that matters, and nothing else in this file
		// looks at that field beyond `!= ""`.
		corpusDigest = "ac4c14ef47e0c42807f47c88abc49da672884864392930fb07ecf4fe71f8fa23"
	)
	corpus := loadOffTargetCorpus(t)
	var offTargets, healthy int
	worstSrc, worstTgt, shortestOff, bestOffTgt := 0.0, 1.0, 1<<30, 0.0
	for _, s := range corpus {
		switch s.Label {
		case "off_target":
			offTargets++
			if s.Letters < shortestOff {
				shortestOff = s.Letters
			}
			if sh := s.targetShare(); sh > bestOffTgt {
				bestOffTgt = sh
			}
		default:
			healthy++
			if sh := s.share(); sh > worstSrc {
				worstSrc = sh
			}
			if sh := s.targetShare(); sh < worstTgt {
				worstTgt = sh
			}
		}
	}
	for _, c := range []struct {
		name        string
		gotN, wantN int
	}{
		{"off-target rows", offTargets, wantOffTargets},
		{"healthy rows", healthy, wantHealthy},
		{"shortest off-target, letters", shortestOff, wantShortestOffChars},
	} {
		if c.gotN != c.wantN {
			t.Errorf("%s = %d, recorded %d — if the corpus was legitimately re-taken, re-record it here in the same commit and say so in the report", c.name, c.gotN, c.wantN)
		}
	}
	for _, c := range []struct {
		name      string
		got, want float64
	}{
		{"worst healthy source-script share", worstSrc, wantWorstHealthySrc},
		{"worst healthy target-script share", worstTgt, wantWorstHealthyTgt},
		{"best off-target target-script share", bestOffTgt, wantBestOffTargetTgt},
	} {
		if diff := c.got - c.want; diff > tol || diff < -tol {
			t.Errorf("%s = %.6f, recorded %.6f — the number the contract is argued from moved; re-record it deliberately, or find out what edited the fixture", c.name, c.got, c.want)
		}
	}
	// The whole population, so that a row which is not an extremum cannot be softened, relabelled, swapped
	// for an easier one under its own id, or duplicated, without saying so.
	if got := offTargetCorpusDigest(corpus); got != corpusDigest {
		t.Errorf("the corpus accounting changed:\n  is:       %s\n  recorded: %s\nAll EIGHT fields of every "+
			"row feed this — id, label, the four counts, provenance and head. If the corpus was legitimately "+
			"re-taken, re-record the digest in the same commit and say so in the report; if it was not, find "+
			"out what edited the fixture.", got, corpusDigest)
	}
}

// offTargetCorpusDigest folds every specimen's identity and accounting — sorted by id, so the digest is a
// property of the SET and not of the file's line order.
func offTargetCorpusDigest(corpus []offTargetSpecimen) string {
	rows := append([]offTargetSpecimen(nil), corpus...)
	sort.Slice(rows, func(i, j int) bool { return rows[i].ID < rows[j].ID })
	h := sha256.New()
	for _, s := range rows {
		fmt.Fprintf(h, "%s|%s|%d|%d|%d|%d|%s|%s\n", s.ID, s.Label, s.SourceScriptRunes, s.TotalRunes,
			s.Letters, s.TargetScriptLetters, s.Provenance, s.Head)
	}
	return hex.EncodeToString(h.Sum(nil))
}

// TestTheLiveEchoGateKeepsItsMeasuredMargin is row 113 itself: the number the shipping gate actually
// achieves, computed by the shipping function, over a corpus that lives in the repository.
//
// ⛔ ITS THRESHOLDS ARE A STOP, not a target. Row 113 says so in as many words: if the fixture stops
// reproducing the recall at zero false positives, or the margin falls under 3×, the premise row 46 is
// built on has fallen and building on it is a waste. They are asserted here so that a later edit to the
// detector, to the threshold or to the preprocessing has to face them.
func TestTheLiveEchoGateKeepsItsMeasuredMargin(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	scripts := lang.LangScripts("zh")
	if len(scripts) == 0 {
		t.Fatal("the source language declares no scripts — the detector would measure nothing (D39.64 П2)")
	}
	var caught, offTargets, healthy, falsePositives int
	worstHealthy := 0.0
	var missed []string
	for _, s := range corpus {
		// The REAL function, over a surrogate carrying the specimen's own rune accounting.
		got := sourceScriptShare(s.surrogate(), scripts)
		if want := s.share(); got != want {
			t.Fatalf("%s: the detector reads %.6f off its own accounting, want %.6f — the surrogate and the fixture disagree", s.ID, got, want)
		}
		switch s.Label {
		case "off_target":
			offTargets++
			if got > cjkEchoThreshold {
				caught++
			} else {
				missed = append(missed, fmt.Sprintf("%s (%s, %d letters, share %.4f)", s.ID, s.Provenance, s.Letters, got))
			}
		default:
			healthy++
			if got > cjkEchoThreshold {
				falsePositives++
			}
			if got > worstHealthy {
				worstHealthy = got
			}
		}
	}
	// ⛔ STOP 1 — zero false positives. A gate that flags healthy Russian is not a margin question at all.
	if falsePositives != 0 {
		t.Fatalf("%d of %d healthy outputs are flagged by the live gate — row 46's premise has fallen", falsePositives, healthy)
	}
	// ⛔ STOP 2 — the recall the contract quotes. Stated as a RATIO so the corpus may grow: the original
	// number was 18 of 20 on a narrower slice of the same stand, and this wider one is 28 of 30.
	const wantRecall = 18.0 / 20.0
	// ⚠ THE DENOMINATOR IS CHECKED FIRST, and not out of tidiness: 0/0 is NaN and every comparison against
	// NaN is false, so a fixture whose off-target class had been emptied would sail through the recall STOP
	// below reporting nothing. A guard whose own arithmetic can excuse it is not a guard.
	if offTargets == 0 {
		t.Fatal("the corpus carries no off-target specimen — the recall STOP below cannot fail and this test would be enforcing nothing")
	}
	if got := float64(caught) / float64(offTargets); got < wantRecall {
		t.Fatalf("the live gate catches %d of %d off-target outputs (%.3f), below the quoted %.3f — the base row 46 must beat has moved. Missed: %s",
			caught, offTargets, got, wantRecall, strings.Join(missed, "; "))
	}
	// ⛔ STOP 3 — the REAL margin, which is the number nobody had recorded: the worst HEALTHY output's
	// source-script share against the threshold that judges it.
	if worstHealthy <= 0 {
		t.Fatal("no healthy specimen carries any source-script rune — the margin would be infinite and meaningless; the corpus must keep the hard cases")
	}
	margin := cjkEchoThreshold / worstHealthy
	if margin < 3 {
		t.Fatalf("the live gate's margin is %.2f× (worst healthy %.6f against threshold %.2f), under the 3× floor row 113 stops at", margin, worstHealthy, cjkEchoThreshold)
	}
	t.Logf("live echo gate: caught %d/%d off-target, %d false positives on %d healthy, worst healthy %.6f, margin %.2f×",
		caught, offTargets, falsePositives, healthy, worstHealthy, margin)
}

// TestTheGateMissesExactlyTheLatinOffTargetClass names WHAT the residual hole is, because «28 of 30» on
// its own invites the reading that two specimens are noise. They are not: the misses are the two fluent
// ENGLISH drafts, and they are missed BY CONSTRUCTION — an output with no source-script rune in it scores
// zero on a source-script detector however wrong it is. That is precisely the hole row 46 exists to close,
// and it is the reason a target-side screen is a different mechanism rather than a better threshold.
func TestTheGateMissesExactlyTheLatinOffTargetClass(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	scripts := lang.LangScripts("zh")
	for _, s := range corpus {
		if s.Label != "off_target" {
			continue
		}
		if sourceScriptShare(s.surrogate(), scripts) > cjkEchoThreshold {
			continue
		}
		// A missed specimen must carry NO source script at all. A miss with source runes in it would be a
		// threshold problem — a different defect with a different cure — and must not hide in this class.
		if s.SourceScriptRunes != 0 {
			t.Errorf("%s carries %d source-script runes and is still missed: that is a THRESHOLD miss, not the Latin class (%s)",
				s.ID, s.SourceScriptRunes, s.Provenance)
		}
	}
}

// TestTheTargetScreenChangesExactlyTheTwoLinesItWasAskedTo is EFFORT_HANDLE §5.4 step 5 executed, and it
// is the acceptance criterion the plan set before the code existed: run the proposed verdict offline over
// the corpus, diff the dispositions, and expect exactly the two English drafts to move. «STOP and escalate
// if a single HEALTHY line changes» — so the healthy half is a Fatal here, not an Error.
//
// It is the whole justification for the screen in one number. The echo gate already catches 28 of 30 at
// zero false positives, so the screen is only worth its snapshot bump if it closes that residue and
// touches nothing else; a screen that also moved healthy lines would be paying for a re-purchase of every
// book with false flags.
func TestTheTargetScreenChangesExactlyTheTwoLinesItWasAskedTo(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	srcScripts, tgtScripts := lang.LangScripts("zh"), lang.LangScripts("ru")
	if len(srcScripts) == 0 || len(tgtScripts) == 0 {
		t.Fatal("both languages must declare their scripts — the rules are data-driven (D39.64 П2)")
	}
	var newlyFlagged, healthyMoved []string
	for _, s := range corpus {
		// What the gate said BEFORE the screen existed: the echo rule alone.
		echoed := sourceScriptShare(s.surrogate(), srcScripts) > cjkEchoThreshold
		// What it says now, for a specimen the echo rule let through. The surrogate is bound to the
		// specimen's own accounting the same way the echo half is, so a fixture edited on one side and not
		// the other cannot make this measurement quietly meaningless.
		scr := langscreen.Screen(s.targetSurrogate(), tgtScripts)
		if scr.Verdict != langscreen.Abstain && scr.Share != s.targetShare() {
			t.Fatalf("%s: the screen reads %.6f off its own accounting, want %.6f", s.ID, scr.Share, s.targetShare())
		}
		screened := !echoed && scr.Verdict == langscreen.OffTarget
		if !screened {
			continue
		}
		line := fmt.Sprintf("%s (%s, %d letters, target share %.4f)", s.ID, s.Provenance, s.Letters, s.targetShare())
		newlyFlagged = append(newlyFlagged, line)
		if s.Label == "healthy" {
			healthyMoved = append(healthyMoved, line)
		}
	}
	// ⛔ THE STOP. A healthy completion newly flagged is a paid unit lost and a book re-purchased for
	// nothing; the plan says escalate rather than proceed.
	if len(healthyMoved) > 0 {
		t.Fatalf("the screen flags %d HEALTHY completion(s) — STOP: %s", len(healthyMoved), strings.Join(healthyMoved, "; "))
	}
	if len(newlyFlagged) != 2 {
		t.Fatalf("the screen must move exactly the two off-target completions the echo rule cannot see, moved %d: %s",
			len(newlyFlagged), strings.Join(newlyFlagged, "; "))
	}
	t.Logf("the target screen newly flags: %s", strings.Join(newlyFlagged, "; "))
}

// TestTheTwoScreensTogetherLeaveNoOffTargetBehind is the pair's joint recall, which is the number that
// actually answers row 46: not «the screen works» but «what still ships silently after both rules run».
func TestTheTwoScreensTogetherLeaveNoOffTargetBehind(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	srcScripts, tgtScripts := lang.LangScripts("zh"), lang.LangScripts("ru")
	var offTargets, caught, falsePositives int
	var escaped []string
	for _, s := range corpus {
		flagged := sourceScriptShare(s.surrogate(), srcScripts) > cjkEchoThreshold ||
			langscreen.Screen(s.targetSurrogate(), tgtScripts).Verdict == langscreen.OffTarget
		switch s.Label {
		case "off_target":
			offTargets++
			if flagged {
				caught++
			} else {
				escaped = append(escaped, fmt.Sprintf("%s (%s)", s.ID, s.Provenance))
			}
		default:
			if flagged {
				falsePositives++
			}
		}
	}
	if falsePositives != 0 {
		t.Fatalf("%d healthy completions are flagged by the pair of rules", falsePositives)
	}
	if caught != offTargets {
		t.Fatalf("%d of %d off-target completions still ship as ok: %s", offTargets-caught, offTargets, strings.Join(escaped, "; "))
	}
	t.Logf("echo gate + target screen: %d/%d off-target caught, %d false positives", caught, offTargets, falsePositives)
}

// TestABankRoleReplyIsExemptFromBothRules is the false-positive class that would arrive WITH this pack and
// not before it. A term table's letters are engine identifiers and source terms, so it scores near zero on
// the target screen and high on the echo rule — every healthy classifier batch would be flagged twice.
// The classifier was already flagged by the echo rule and it was invisible because no shipping config ran
// it; backlog row 140 runs it.
func TestABankRoleReplyIsExemptFromBothRules(t *testing.T) {
	table := strings.Repeat("元石\tterm\tnone\n方源\tname\tmale\n", 40)
	in := classifyInput{
		Output: table, Finish: llm.FinishStop, TargetLang: "ru",
		SourceScripts: lang.LangScripts("zh"), TargetScripts: lang.LangScripts("ru"),
	}
	if got := classify(in); got.Reason == reasonOK {
		t.Fatal("premise broken: without the exemption this reply must be flagged, or the test proves nothing")
	}
	in.NonProseReply = true
	if got := classify(in); got.Reason != reasonOK {
		t.Fatalf("a bank role's table must pass both rules, got %q (%s)", got.Reason, got.Detail)
	}
	// And the exemption must be keyed on the role SET, not on one role: the classifier's reply has the
	// same shape as the terminologist's and was outside it.
	for _, role := range []string{roleTerminologist, roleClassifier} {
		if !isBankRole(role) {
			t.Errorf("%s answers a term table and must be exempt", role)
		}
	}
	if isBankRole(roleTranslator) || isBankRole(roleEditor) {
		t.Error("a prose role must NOT be exempt — the exemption is what makes the rules meaningless")
	}
}

// TestTheScreenConstantsStillSitWhereTheCorpusPutThem binds langscreen's two calibrated numbers to the
// corpus they were taken from. They live in that package as bare constants with the measurement written in
// prose beside them, and prose does not fail: a fixture edited until the gap closed, or a floor raised past
// the healthy minimum, would leave every test in that package green because they compare the constants with
// constants. Here they are compared with the data.
func TestTheScreenConstantsStillSitWhereTheCorpusPutThem(t *testing.T) {
	corpus := loadOffTargetCorpus(t)
	worstHealthy, bestOffTarget, shortestOffTarget := 1.0, 0.0, 1<<30
	for _, s := range corpus {
		switch s.Label {
		case "off_target":
			if sh := s.targetShare(); sh > bestOffTarget {
				bestOffTarget = sh
			}
			if s.Letters < shortestOffTarget {
				shortestOffTarget = s.Letters
			}
		default:
			if sh := s.targetShare(); sh < worstHealthy {
				worstHealthy = sh
			}
		}
	}
	if !(bestOffTarget < langscreen.OnTargetFloor && langscreen.OnTargetFloor < worstHealthy) {
		t.Fatalf("the floor %.4f no longer sits in the corpus's empty interval [%.4f, %.4f]", langscreen.OnTargetFloor, bestOffTarget, worstHealthy)
	}
	// ⛔ The evidence floor may not rise past the SHORTEST off-target completion: above it, the screen would
	// abstain on a real event, which is the one thing an abstention must never hide.
	if langscreen.MinLetters > shortestOffTarget {
		t.Fatalf("the evidence floor %d is above the shortest off-target completion (%d letters) — a real event would be excused as too short", langscreen.MinLetters, shortestOffTarget)
	}
}