The zh pair's other direction, and a rule pack that mostly says no

`pair_lang` had always been answering a second question nobody asked: it
says which two languages, and every surface built on it assumed English
was the one being learned. That is why hanzi is never tokenized, never
spell-checked, never glossed — correct for a Mandarin native practising
English, backwards for an English native practising Mandarin.
`users.direction` (migration 0016) separates the two questions; a
`zh-learner` pair code would have been cheaper and would have made two
directions of one pair look like two unrelated languages to every query.

Segmentation is what replaces `wordAt` where there are no spaces: a
shortest-path walk over log-probabilities, 232 ms and 14 MB for 188,522
words. The browser gets the word list because segmentation runs on hover;
the server keeps the whole dictionary. Their coverage gates come out
opposite on purpose — the client list is frequency-gated because the
segmentation is measurably identical without the tail, and the dictionary
is gated by nothing, because its only power is to explain and the word a
learner stops on is the rare one.

The 错别字 pack is 24 confusable pairs behind two mechanical gates. One
admits a pair only if the wrong form is not a dictionary word and the
right form is, which is why it refuses 自已 for 自己 — a real error whose
wrong form is a headword. The other asks the segmenter whether the two
characters already belong to two different words, without which 自己经常,
睡觉的时候 and 不知到底 would all be corrupted silently into text still
made of real characters.

Not deployed (this carries a migration), not seen in a browser, and no
account has ever been in the learner direction. The IME composition
guards were in scope and are not done — see BUILD_PLAN Phase 26.

internal/lexicon/data.go

@@ -36,3 +36,13 @@ var glossGz []byte
 //
 //go:embed data/phonetic.json.gz
 var phoneticGz []byte
+
+// hanziGz is the gzipped Chinese→English map: simplified headword → [[pinyin,
+// senses], …]. Built from CC-CEDICT (scripts/build_cedict.py), unfiltered — the
+// word a learner stops on is the one they do not know, so this is the one
+// dataset here with no frequency gate. Loaded on its own sync.Once (see
+// hanzi.go), not with the four above, because only a learner-direction account
+// ever asks for it.
+//
+//go:embed data/hanzi.json.gz
+var hanziGz []byte

internal/lexicon/handlers.go

@@ -42,6 +42,36 @@ func (h *Handler) GlossRoutes() chi.Router {
 	return r
 }
 
+// HanziRoutes returns the router mounted at /api/hanzi — a Chinese word to its
+// pinyin and English senses, for a writer going the other way through the zh
+// pair (`users.direction = 'learning_pair'`).
+//
+// It does not go through [Handler.providerFor], and that is not an oversight.
+// providerFor picks a dictionary by the writer's *pair*, to answer "what does
+// this English word mean in her language" — a question whose answer differs per
+// pair. This endpoint asks the opposite question of exactly one language, and
+// [auth.SupportsLearnerDirection] already guarantees that language is Chinese.
+// Routing it through the pair would add a database read per hover to choose
+// between one option and itself.
+func (h *Handler) HanziRoutes() chi.Router {
+	r := chi.NewRouter()
+	r.Get("/{word}", h.hanzi)
+	return r
+}
+
+// hanzi answers a Chinese word lookup. Like the other two, a miss is a 200 with
+// empty lists — a hover that lands on a word the dictionary has never heard of
+// is an ordinary thing to happen while reading, and the tooltip simply doesn't
+// open.
+func (h *Handler) hanzi(w http.ResponseWriter, r *http.Request) {
+	res, err := h.Set.Hanzi(pathWord(r))
+	if err != nil {
+		writeLookupErr(w, err)
+		return
+	}
+	writeLookup(w, res)
+}
+
 // providerFor returns the provider for the caller's language pair.
 //
 // The pair language is read here rather than threaded down because a word

internal/lexicon/hanzi.go

@@ -0,0 +1,144 @@
+package lexicon
+
+import (
+	"fmt"
+	"strings"
+	"sync"
+	"unicode"
+)
+
+// The Chinese half of the lexicon: a word written in hanzi to its pinyin and
+// English senses. This is the mirror image of `gloss` — that one reads English
+// and answers in Chinese, for a Mandarin native practising English; this one
+// reads Chinese and answers in English, for the other direction of the same
+// pair (`users.direction = 'learning_pair'`).
+//
+// It is deliberately not folded into [Lexicon.load]. That method reads four
+// datasets on the first lookup of any kind, and this one is 3.1 MB gzipped that
+// only a learner-direction account will ever ask for — every other writer would
+// pay the decompression and the resident memory for a map they never touch. Its
+// own sync.Once means the cost lands on the first Chinese hover and nowhere
+// else.
+
+// HanziReading is one pronunciation of a word and the senses it carries in that
+// pronunciation. A word usually has one; the ones that have two are why this is
+// a list rather than a pair of strings. 得 is dé, "to obtain", *and* de, the
+// particle that makes 说得很好 mean "speaks well" — a learner shown only the
+// first has been told something false about the sentence in front of them.
+type HanziReading struct {
+	Pinyin string `json:"pinyin"`
+	Senses string `json:"senses"`
+}
+
+// HanziChar is one character of a word that the dictionary could not answer as
+// a whole. See [Lexicon.Hanzi].
+type HanziChar struct {
+	Char   string `json:"char"`
+	Pinyin string `json:"pinyin"`
+	Senses string `json:"senses"`
+}
+
+// HanziResult is what a Chinese word lookup answers. Readings is empty for a
+// word the dictionary does not have, in which case Chars may carry the
+// character-by-character reading instead.
+type HanziResult struct {
+	Word     string         `json:"word"`
+	Readings []HanziReading `json:"readings"`
+	Chars    []HanziChar    `json:"chars"`
+}
+
+type hanziStore struct {
+	once sync.Once
+	err  error
+	// word → [[pinyin, senses], …], exactly as scripts/build_cedict.py writes it.
+	entries map[string][][]string
+}
+
+var hanzi hanziStore
+
+func (h *hanziStore) load() {
+	h.once.Do(func() {
+		if err := gunzipJSON(hanziGz, &h.entries); err != nil {
+			h.err = fmt.Errorf("load hanzi: %w", err)
+		}
+	})
+}
+
+// maxHanziChars caps the per-character fallback. A run longer than this is
+// almost certainly a phrase the segmenter split badly rather than a word, and
+// spelling out eight characters one at a time is a wall, not a hint.
+const maxHanziChars = 6
+
+// Hanzi returns the pinyin and English senses of a Chinese word.
+//
+// There is no de-inflection walk here, and its absence is a fact about the
+// language rather than an omission: Chinese words do not inflect, so the
+// candidate forms [lookupGloss] tries for "running" → "run" have no analogue.
+// A lookup either hits the headword or it does not.
+//
+// What it does instead is fall back to the characters. The segmentation word
+// list is a superset of this dictionary — every glossable word can be
+// segmented, but jieba knows ordinary compounds CC-CEDICT has no entry for — so
+// a hover really can land on a word with nothing to say about it. Chinese
+// compounds are usually transparent from their parts (电脑 is "electric brain"),
+// which makes the character reading a genuinely useful second answer rather
+// than a consolation prize. It is returned as its own field so the surface can
+// say which of the two it is showing; a caller that only wants whole words can
+// ignore it.
+func (l *Lexicon) Hanzi(word string) (HanziResult, error) {
+	hanzi.load()
+	if hanzi.err != nil {
+		return HanziResult{}, hanzi.err
+	}
+
+	norm := strings.TrimSpace(word)
+	res := HanziResult{Word: word, Readings: []HanziReading{}, Chars: []HanziChar{}}
+	if norm == "" {
+		return res, nil
+	}
+
+	if rows, ok := hanzi.entries[norm]; ok {
+		res.Readings = toReadings(rows)
+		return res, nil
+	}
+
+	chars := []rune(norm)
+	if len(chars) < 2 || len(chars) > maxHanziChars {
+		// A single character that missed has no parts to fall back to, and a long
+		// run is not a word. Either way the honest answer is nothing.
+		return res, nil
+	}
+	for _, r := range chars {
+		if !unicode.Is(unicode.Han, r) {
+			// Mixed input (a stray letter or digit inside the run) is not something
+			// the character reading can explain, and guessing at the hanzi parts of
+			// it would be worse than silence.
+			return HanziResult{Word: word, Readings: []HanziReading{}, Chars: []HanziChar{}}, nil
+		}
+		rows, ok := hanzi.entries[string(r)]
+		if !ok {
+			continue
+		}
+		first := toReadings(rows)
+		if len(first) == 0 {
+			continue
+		}
+		res.Chars = append(res.Chars, HanziChar{
+			Char:   string(r),
+			Pinyin: first[0].Pinyin,
+			Senses: first[0].Senses,
+		})
+	}
+	return res, nil
+}
+
+func toReadings(rows [][]string) []HanziReading {
+	out := make([]HanziReading, 0, len(rows))
+	for _, row := range rows {
+		if len(row) < 2 {
+			continue
+		}
+		out = append(out, HanziReading{Pinyin: row[0], Senses: row[1]})
+	}
+	return out
+}

internal/lexicon/hanzi_test.go

@@ -0,0 +1,150 @@
+package lexicon
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+)
+
+// The Chinese direction of the lexicon, against the real embedded asset — not a
+// fixture. The dataset is built by scripts/build_cedict.py, which asserts its
+// own invariants at build time; what these assert is that the *lookup* over it
+// behaves, including on the entries the build script goes out of its way to keep.
+
+func TestHanziLookup(t *testing.T) {
+	l := New()
+
+	res, err := l.Hanzi("公园")
+	if err != nil {
+		t.Fatalf("lookup 公园: %v", err)
+	}
+	if len(res.Readings) == 0 {
+		t.Fatal("公园 has no readings")
+	}
+	// Tone marks, not the numbered pinyin CC-CEDICT stores. The number is the
+	// storage format; the marks are what a learner reads.
+	if got := res.Readings[0].Pinyin; got != "gōngyuán" {
+		t.Errorf("公园 pinyin = %q, want gōngyuán", got)
+	}
+	if !strings.Contains(res.Readings[0].Senses, "park") {
+		t.Errorf("公园 senses = %q, want something about a park", res.Readings[0].Senses)
+	}
+	// A word answered whole says nothing about its characters — the fallback is
+	// the other branch, and sending both would double the payload of the common
+	// case to no purpose.
+	if len(res.Chars) != 0 {
+		t.Errorf("a whole-word hit also returned %d characters", len(res.Chars))
+	}
+}
+
+// 得 is the reason readings are a list. Answered with only dé "to obtain", a
+// learner hovering it in 说得很好 has been told something false about the
+// sentence they are looking at.
+func TestHanziParticleCarriesItsGrammaticalReading(t *testing.T) {
+	l := New()
+
+	for _, particle := range []string{"的", "地", "得"} {
+		res, err := l.Hanzi(particle)
+		if err != nil {
+			t.Fatalf("lookup %s: %v", particle, err)
+		}
+		var found bool
+		for _, r := range res.Readings {
+			if r.Pinyin == "de" {
+				found = true
+			}
+		}
+		if !found {
+			t.Errorf("%s never reads as neutral \"de\": %+v", particle, res.Readings)
+		}
+	}
+}
+
+// The fallback the segmentation gap makes necessary: jieba knows ordinary
+// compounds CC-CEDICT has no headword for, so a hover can land on a real word
+// with no entry. Chinese compounds are usually transparent from their parts, so
+// the characters are a real second answer.
+func TestHanziFallsBackToCharacters(t *testing.T) {
+	l := New()
+
+	// Constructed rather than borrowed from the corpus: a word that CC-CEDICT
+	// *does* carry would test the other branch, and which compounds it happens to
+	// omit is not something a test should pin.
+	const made = "猫书"
+	if _, ok := hanzi.entries[made]; ok {
+		t.Skipf("%s has become a real headword; pick another compound", made)
+	}
+	res, err := l.Hanzi(made)
+	if err != nil {
+		t.Fatalf("lookup %s: %v", made, err)
+	}
+	if len(res.Readings) != 0 {
+		t.Fatalf("%s answered as a whole word: %+v", made, res.Readings)
+	}
+	if len(res.Chars) != 2 {
+		t.Fatalf("character fallback gave %d entries, want 2: %+v", len(res.Chars), res.Chars)
+	}
+	if res.Chars[0].Char != "猫" || !strings.Contains(res.Chars[0].Senses, "cat") {
+		t.Errorf("first character = %+v, want 猫 ~ cat", res.Chars[0])
+	}
+	if res.Chars[0].Pinyin != "māo" {
+		t.Errorf("猫 pinyin = %q, want māo", res.Chars[0].Pinyin)
+	}
+}
+
+func TestHanziMisses(t *testing.T) {
+	l := New()
+
+	for name, word := range map[string]string{
+		// A single character with no entry has no parts to fall back to.
+		"lone unknown character": "龥",
+		"empty":                  "",
+		"whitespace":             "  ",
+		// Not Chinese at all: the English tokenizer owns these, and answering
+		// would mean guessing.
+		"english": "hello",
+		"mixed":   "猫cat",
+		// Longer than a word: a bad segmentation, not something to spell out
+		// character by character.
+		"a whole clause": "我今天早上去公园跑步了",
+	} {
+		res, err := l.Hanzi(word)
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		if len(res.Readings) != 0 || len(res.Chars) != 0 {
+			t.Errorf("%s (%q) answered with %+v / %+v", name, word, res.Readings, res.Chars)
+		}
+	}
+}
+
+func TestHanziEndpoint(t *testing.T) {
+	h := NewHandler(nil, NewSet(nil))
+	r := chi.NewRouter()
+	r.Mount("/hanzi", h.HanziRoutes())
+
+	w := httptest.NewRecorder()
+	r.ServeHTTP(w, httptest.NewRequest(http.MethodGet, "/hanzi/"+"跑步", nil))
+	if w.Code != http.StatusOK {
+		t.Fatalf("status = %d", w.Code)
+	}
+	var got HanziResult
+	if err := json.Unmarshal(w.Body.Bytes(), &got); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if got.Word != "跑步" || len(got.Readings) == 0 || got.Readings[0].Pinyin != "pǎobù" {
+		t.Fatalf("response = %+v", got)
+	}
+
+	// A miss is a 200 with empty lists, like the other two lookups — the tooltip
+	// quietly doesn't open rather than showing an error over her writing.
+	w = httptest.NewRecorder()
+	r.ServeHTTP(w, httptest.NewRequest(http.MethodGet, "/hanzi/zzz", nil))
+	if w.Code != http.StatusOK {
+		t.Fatalf("miss: status = %d, want 200", w.Code)
+	}
+}

internal/lexicon/provider.go

@@ -109,3 +109,13 @@ func (g glossless) Lookup(word string) (Result, error) {
 func (g glossless) Gloss(word string) (GlossResult, error) {
 	return GlossResult{Word: word}, nil
 }
+
+// Hanzi answers a Chinese-word lookup from the embedded CC-CEDICT map.
+//
+// It is on the Set rather than on [Provider] because it is not the same
+// question the other two ask. Lookup and Gloss vary by pair — which is why they
+// are behind an interface with two implementations — while this one is asked of
+// Chinese or not at all: the learner direction exists for exactly one pair (see
+// auth.learnerPairs), and DreamDict's own CC-CEDICT would be a second copy of
+// the same dictionary, chosen by a rule with one branch.
+func (s *Set) Hanzi(word string) (HanziResult, error) { return s.embedded.Hanzi(word) }