feat: voice and language capture through existing fingerprint layers

[daveh-beep]

Summary

Documents how voice and language flow through Ghost's existing fingerprint model — no new domain, no schema change, no new dimensions. Supersedes #49.

Why not a new domain

#49 proposed communication voice as a parallel dimension set (tone/structure/register/boundaries with 0–1 scalars and a 20-dim embedding). That design predates the fingerprint-package model and has aged out on three axes:

1. Targets a removed package. Files land in packages/ghost-drift/, which was folded into packages/ghost. The frontmatter/body partition rules and ghost-drift verbs it builds on are legacy.
2. The current model already absorbed the concern. prose.summary.tone exists, composition.yml has kind: content patterns, and enforcement/checks.yml ships forbidden-regex/required-regex detectors — which is what the proposed boundaries block reinvents.
3. Scored voice scalars are a human call, not a measurement. Curated prose rules with evidence, owned through Git review, fit Ghost's model; an embedded voice ontology does not. Ghost stays the deterministic calculator and never ships anyone's style guide.

What this adds

• docs/language-fingerprints.md — the mapping: voice intent → prose.yml (tone words, principles, experience contracts for exact-wording surfaces); copy material and external writing standards → inventory.yml (sources entries point at maintained standards instead of forking them); copy shapes → composition.yml kind: content patterns; the mechanically detectable subset → enforcement/checks.yml.
• Normativity-level mapping (RFC 2119 spirit): a standard's must → status: active check (blocks); should → status: proposed (advisory in ghost review); may/contextual → composition guidance only, never a gate.
• references/voice.md — skill recipe: inventory user-facing strings → read the declared standards source → draft smallest evidence-backed entries → promote only the deterministic subset to checks → human curates, Git approves.
• Registered in SKILL.md, install/manifest.json, README resources, plus a minor changeset.

All examples are org-agnostic per the repo's own oss-language-is-portable principle: any project with a style guide wires it in via inventory.sources; organization-specific voice content lives in each adopting repo's own fingerprint, not in Ghost.

Test plan

• pnpm build clean
• pnpm test — 338/338 passing (install-manifest test covers voice.md)
• pnpm check guards pass (terminology, file-size, install-bundle, CLI manifest in sync)
• No runtime code changes; docs + skill bundle + manifest only