refactor(build-mcpb): split language-specific content into per-language reference files

[JoeCardoso13]

Supersedes #57.

Problem

build-mcpb carries two complete language tracks (Python and TypeScript) interleaved throughout every phase and reference file. Every invocation picks a language once in Phase 0 and never changes it — meaning half the content loaded into context at every step is irrelevant dead weight.

Splitting into two separate skills (as proposed in #57) was considered and rejected — we don't want two skills to maintain.

Proposed Fix

Keep build-mcpb as a single skill but separate the language-specific content into per-language reference files. The workflow phases become language-agnostic, pointing to {lang}-specific references using the language established in Phase 0.

Reference files to split

Current	Becomes
references/PATTERNS.md	references/PATTERNS-PY.md + references/PATTERNS-TS.md
references/CONVENTIONS.md	references/CONVENTIONS-PY.md + references/CONVENTIONS-TS.md
references/SKILL_FORMAT.md	references/SKILL_FORMAT-PY.md + references/SKILL_FORMAT-TS.md

Workflow phases to simplify (remove all "If Python... / If TypeScript..." branching)

• phase-0-bootstrap.md — dual template substitution lists
• phase-2-scaffold.md — dual file existence checklists
• phase-3-implement-and-verify.md — dual implementation orders and verify commands
• phase-4-validate-bundle.md — dual runtime validation commands
• phase-5-embed-skill.md — dual wiring instructions

Phases reference the appropriate file with e.g. "See references/PATTERNS-{lang}.md" where {lang} is PY or TS, locked in at Phase 0.

Outcome

• Agent only loads one language's reference files per run
• Workflow phases are lean and language-agnostic
• Single skill, single pipeline — no additional maintenance burden