#51 follow-up: prompt change (layer A) never landed in generate-instructions.md

[jstet]

Follow-up to #51.

The PR/commit message for #51 describes three layers. Layer A (prompt) isn't in the commit. 3e398e3 only touches lead.ts, lead.test.ts, the e2e snapshot and one fixture. scripts/skill-content/generate-instructions.md:35 still has the old wording:

- Use `"Sonstiges"` + a follow-up `text` question with `relevant` logic instead of `or_other`

The hard rule described in the PR is missing: follow-up is required, named <parent>_sonstiges / <parent>_other, selected(${parent}, '<code>') vs ${parent} = '<code>', and the same for yes/no follow-ups. The generator gets no new guidance, so every broken follow-up has to be caught by the repair loop, and that loop has holes (see the related issues).

Minor: the new e2e fixture missing-relevant-on-follow-up only adds a snapshot of [] ("passes the formtransform validator"). It doesn't record that relevant is missing, so it doesn't document the failure mode the way the PR says it does.

Fix: apply the prompt change from the PR description to generate-instructions.md. Rebuild whatever is generated from it, and check that npm run test:e2e snapshots still match. Optionally, make the fixture assert that looksLikeFollowUp flags it.

[jstet]

More context from starting on this:

• The naming in the PR's rule clashes with the prompt. generate-instructions.md:12 says names have "no underscores", but the rule described in wizard: catch and repair Sonstiges follow-ups without relevant (#47 follow-up) #51 asks for <parent>_sonstiges. And sanitizeSurvey turns _sonstiges into sonstiges anyway (see looksLikeFollowUp stops detecting follow-ups after sanitize; repair loop can hide the bug #53).
• The registry already has a convention for this, and the sanitizer supports it. cdl-survey-types/references/xlsform-syntax.md:25: add a choice with code other, then a companion text row named <question>_other with relevant = ${<question>} = 'other'. sanitize.ts keeps the _other suffix on purpose. The prompt should use exactly this (choice other, label "Sonstiges", companion <question>_other, selected(${q}, 'other') for select_multiple), and say _other is the only allowed underscore.
• The output-format table has no relevant field. question_parser.ts accepts relevant, but the table in generate-instructions.md never tells the model it can output one. Add a row.
• For yes/no follow-ups: place them directly after the parent and set relevant to ${parent} = '<code>'.
• skills/xlsform/generate-instructions.md is the copy the app imports (survey_generator.ts:5). Update it together with scripts/skill-content/, or rebuild it with npm run build:skill.