feat(mentat): improve skill generation prompt

[bketelsen]

Summary

Rewrites buildPrompt() in the mentat generator based on Anthropic's skill authoring best practices, and simplifies normaliseContent().

Changes

buildPrompt() — ~50% shorter, example-driven

• Before: verbose, listed 5 section names explicitly, gave lengthy instructions the model ignores
• After: concise task description + one concrete one-shot example (queue domain); trusts the model to choose structure
• Domain context condensed to a single dense header line
• Frontmatter requirement stated clearly with the domain name pre-filled
• Explicit instruction: only include what the agent can't infer from reading the code

normaliseContent() — simplified

• Removed the fragile \n---\n heuristic that could silently truncate mid-document content (e.g. when a horizontal rule appears in the body)
• Retained: code-fence stripping, trailing stats-line stripping, single-newline normalisation
• Frontmatter injection still handled by generateWith (unchanged)

New tests: generator_internal_test.go

14 unit tests covering:

• All prompt properties (domain fields, frontmatter template, example, no rigid section headers, file sample inclusion, empty-sample guard, multi-language formatting)
• All normaliser edge cases (code fences, stats lines, mid-document HR regression, empty input, trailing newline invariant)

Verification

just test   # all packages pass
just build-mentat  # exit 0