Church Translate, as it runs at the campuses

Live sermon translation into 100+ languages, running entirely on one Mac per
campus: audio in, voice detection, Whisper on the GPU, NLLB-200 in its own
process, and a built-in web server the congregation reads from on their
phones. No audio leaves the building.

This is the source only. What used to sit beside it — 14 GB of speech and
translation models, 8 GB of voice bundles, 6 GB of natural voices, the Electron
caches, and the app data holding the operator PIN and every service transcript
— is machine state and downloaded assets, and stays in the NAS backup where it
belongs. The vendored node, cloudflared and Windows ffmpeg binaries are out too
(533 MB, re-fetched for packaging); the browser-side runtime beside them is
kept, so a checkout still serves the congregation page.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/PLAN-spanish-variants.md

@@ -0,0 +1,90 @@
+# Plan — three Spanish variants for live preaching
+
+## The ask
+Operator, relaying a native Spanish speaker: the Spanish output is Latin
+American, and three variants are wanted — **Latin American, Spain, Mexican** —
+"fine tuned exactly to how it's supposed to be". No native speaker will be
+provided; research from the web instead. Nothing may be left out.
+
+## Finding 1 — the taxonomy overlaps
+es-MX is INSIDE es-419, not beside it. Their differences are lexical only
+(*computadora/ordenador*, *carro/coche*, *ícono/icono*). es-ES is the one that
+differs GRAMMATICALLY. So the three deliverables are not three equal problems:
+
+| variant | differs from es-419 by | difficulty |
+|---|---|---|
+| es_419 Latin American | — (this is today's output) | none |
+| es_MX Mexican | word choice | low, deterministic |
+| es_ES Spain | **the whole 2nd-person-plural system** | high |
+
+Reference editions, so each variant has an authority rather than my preference:
+- **es_ES** → Reina-Valera 1960 (uses `vosotros`) / LBLA
+- **es_419** → NBLA (explicitly uses `ustedes` for Latin America) / NVI
+- **es_MX** → NBLA + Mexican lexicon
+
+## Finding 2 — why es-ES cannot be a find-and-replace
+Spanish 3rd-person-plural is IDENTICAL for "ustedes" (you all) and "ellos"
+(they). `hablan` is both. So a target-side rule `hablan → habláis` corrupts
+every genuine third-person sentence: "they were there" becomes "you all were
+there". In scripture. Silently. This is the exact failure class the whole
+fidelity effort exists to prevent, so a naive transform is not shippable.
+
+## Finding 3 — the English source resolves the ambiguity
+This is the key idea and it may remove the need for any model change.
+
+The ambiguity is unresolvable in the Spanish. It is already resolved in the
+ENGLISH, which we still have:
+
+    "Open YOUR Bibles."        -> 2nd person -> ustedes-form -> convert
+    "THEY were there."         -> 3rd person -> ellos-form   -> leave alone
+
+So: convert to `vosotros` ONLY when the English sentence contains 2nd-person
+address AND contains NO 3rd-person-plural marker. Mixed sentences ("tell them
+that you are loved") are left in es-419 form — which a Spaniard understands
+perfectly, it just reads Latin American. **Conservative by construction: the
+failure mode is "not converted", never "wrongly converted".**
+
+## Design
+
+    NLLB -> guard -> cache -> [es-419 base]
+                                 |-- es_MX : lexical layer
+                                 |-- es_ES : gated vosotros transform
+                                 `-- es_419: as-is
+
+Same shape as `pt-variants.js` (por_Latn / por_PT share one NLLB result), and it
+reuses `NLLB_ALIAS` so the extra variants cost no extra inference.
+
+New pseudo-codes `spa_MX` and `spa_ES`, both aliased to `spa_Latn` at the MT
+boundary only — exactly how `por_PT` already works.
+
+### es_ES transform components
+1. **Gate** (on the English source): 2nd-person markers present, 3rd-person-plural
+   markers absent.
+2. **Pronouns/possessives**: ustedes→vosotros, su(s)→vuestro(s)/vuestra(s),
+   les→os, se (reflexive 2pl)→os, a ustedes→a vosotros.
+3. **Verb morphology**: 3pl→2pl. `-an→-áis`; `-en→-éis` for -er verbs but
+   `-en→-ís` for -ir verbs, so the verb's class must be known — a table, not a
+   rule. Irregulars (ser, ir, tener, hacer, poder, querer, decir, venir, saber,
+   dar, ver, poner, salir, traer, oír, haber) need explicit entries.
+4. **Imperatives**: `abran→abrid`, `vengan→venid`, `miren→mirad` — highest
+   frequency in preaching and the most audible difference.
+
+### es_MX layer
+Lexical only. Word-level, no morphology, no gate needed.
+
+## What must not regress
+- Latency: no extra inference. All layers are string work over a shared result.
+- `applyOverrides` still runs last so an operator pin outranks everything.
+- The English caption and archive keep what was actually said.
+- Voice: `speech-text.js` has Spanish book-name rules (`ES_BOOKS`) that fire on
+  `1 Corintios` → `Primera de Corintios`. New variants must not produce forms
+  outside that list or the voice reads "uno Juan".
+- Each variant needs its own voice choice — es_ES should get a Castilian voice
+  (miro/sharvard, measured male), es_MX the Mexican `ald`.
+
+## Open questions for the audit
+1. Is the source-side gate actually safe? Where does it misfire?
+2. Verb-table coverage: what fraction of real sermon Spanish is reachable?
+3. Does adding two pseudo-languages break the shared-cache invariant?
+4. Is `vosotros` even wanted? RVR1960 uses it and is read across Latin America.
+5. Does the listener page cope with three Spanish entries in the picker?