diplomail (Stage D): language detection + lazy translation cache

Replaces the LangUndetermined placeholder with whatlanggo-backed
body detection on every send path, then adds a translation cache
keyed on (message_id, target_lang) populated lazily on the
per-message read endpoint. The noop translator that ships with
Stage D returns engine="noop", which the service treats as
"translation unavailable" — wiring a real backend (LibreTranslate
HTTP client is the documented next step) is a one-file swap.

GetMessage and ListInbox now accept a targetLang argument; the HTTP
layer resolves the caller's accounts.preferred_language and
forwards it. Inbox uses the cache only (never calls the
translator) so bulk reads stay fast under future SaaS backends.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/internal/diplomail/detector/detector.go

@@ -0,0 +1,79 @@
+// Package detector wraps the body-language detection used by the
+// diplomail subsystem. The package exposes a narrow `LanguageDetector`
+// interface so the implementation can be swapped without touching the
+// callers; the default backed-by-whatlanggo detector handles 84
+// natural languages and ships with the embedded statistical profiles.
+//
+// Detection happens only on the body. Subjects are short and
+// frequently template-like ("Re: ..."), so detecting on them adds
+// noise. The diplomail Service feeds the body, captures the BCP 47
+// tag returned here, and stores it in `diplomail_messages.body_lang`.
+package detector
+
+import (
+	"strings"
+	"unicode/utf8"
+
+	"github.com/abadojack/whatlanggo"
+)
+
+// Undetermined is the BCP 47 placeholder stored when detection cannot
+// confidently identify a language (empty body, too-short body, mixed
+// scripts the detector refuses to bet on).
+const Undetermined = "und"
+
+// LanguageDetector is the read-only surface diplomail consumes when
+// it needs to label a message body. Detect must never panic and
+// must never return an error: detection failure simply yields
+// `Undetermined`.
+type LanguageDetector interface {
+	Detect(body string) string
+}
+
+// New returns the package-default detector backed by `whatlanggo`.
+// The instance is safe for concurrent use; whatlanggo's `Detect`
+// reads the embedded profiles without state mutation. Callers that
+// want a fixed allow-list can build their own implementation around
+// the same interface.
+func New() LanguageDetector {
+	return &whatlangDetector{}
+}
+
+type whatlangDetector struct{}
+
+// minRunes is the lower bound on body length below which whatlanggo
+// can flip between near-synonyms; for shorter bodies we return
+// `Undetermined` and let the noop translator skip the slot. The
+// value matches whatlanggo's documented "stable above ~25 runes"
+// guidance.
+const minRunes = 25
+
+// Detect returns the BCP 47 tag for body, or `Undetermined` when the
+// body is empty / too short / whatlanggo refuses to label it. The
+// trim is applied so leading whitespace does not bias the script
+// detector toward Latin. We deliberately do not gate on
+// `info.IsReliable()` because the gate is too conservative for the
+// short sentences typical of in-game mail; a misclassification only
+// hurts the translation cache key, never correctness.
+func (d *whatlangDetector) Detect(body string) string {
+	body = strings.TrimSpace(body)
+	if body == "" {
+		return Undetermined
+	}
+	if utf8.RuneCountInString(body) < minRunes {
+		return Undetermined
+	}
+	info := whatlanggo.Detect(body)
+	tag := info.Lang.Iso6391()
+	if tag == "" {
+		return Undetermined
+	}
+	return tag
+}
+
+// NoopDetector returns the placeholder unconditionally. Used by
+// tests and by Stage A code paths that predate the real detector.
+type NoopDetector struct{}
+
+// Detect always returns `Undetermined` regardless of input.
+func (NoopDetector) Detect(string) string { return Undetermined }