Swiss spelling by string transform, not by prompt

[trick77]

The Swiss orthography prompt note is gone: measured useless on 2026-09-16, gpt-5.4 writes "größer" and "fuer" regardless of wording. German prose is now respelled by a string function on the way out (backend/internal/ask/swiss.go).

• ß to ss always; ae/oe/ue to the umlaut, gated: ue after a/e/q/ä stays (neue, Steuer, Quelle), -uell suffix stays (aktuell) but Müller/befüllen get their ü, a word-final ue is English (Value, Continue), short stem list for names, borrowings and Latin -uen words.
• Never touched: fences, inline spans, identifier-shaped words (camelCase, snake_case, digits, paths, dotted names), diagram node labels (known gap).
• Stream: the renumberer gets a spell hook and holds a word back until its end arrives, so a token boundary inside "Gesch|aeft" still spells right and the reader sees exactly what is stored.
• Title, card names and follow-ups run through the same function via spellFor(lang).
• Eval: the digraphs column now reads Answer.Respelled (what the model wrote wrong), so the deployment's German spelling stays measurable; germanDigraphs in the eval is gone.
• Known false-positive class accepted: compound seams (Kontoerstellung), names not on the list. On MiMo Pro the model writes 0 digraphs per run, on gpt-5.4 about 23.

Verified in the running app with a fake streaming upstream: "Der Gesch" + "aeftsprozess ist gr" + "ößer" arrived as "Der Geschäftsprozess ist grösser", pruefeGroesse and the fence body untouched, follow-up pill "Grösse der Strasse prüfen?", stored answer identical to the stream.