feat: add /swarm-builder:design interview command (#30)

[fdaviddpt]

Adds /swarm-builder:design -- a prose-only interview command that walks a user from nothing to a swarm.json proposal, per issue #30's shape. Ordered after #28 and #29, both already shipped (changelog.d/28.added.md, changelog.d/29.added.md).

What it does

1. Interview. One question at a time, in the user's own domain -- never in agent names -- until the graph is decidable: what runs, what triggers it, which actions are irreversible and who may take them, where a human must stay in the loop, what a run may cost.
2. Restatement. "Here is what I think you need," in the user's own words. They correct it -- the cheapest place to catch wrong-but-plausible, before anything is written.
3. Proposal. The fewest agents that satisfy the need, as swarm.json -- structure only. Never starts from examples/claude-oss.swarm.json or any preset. cost_tokens is the one schema-optional catalogue field and is never guessed, left absent instead; every other required field (summary, model, tools, authority, lifetime, inherits, budget_bytes, context, scripts) has to be settled by the interview.
4. Picture. Renders via /swarm-builder:read, loops step 3-4 until the user confirms the picture is right.
5. Hand off. /swarm-builder:compile creates the MD skeletons once the picture is confirmed.

Review

Two self-review rounds (concurrent Explore + oss:auditor spawns each), seven findings total, all fixed:

• schema-invalid "leave any field absent" guidance narrowed to cost_tokens only
• a stale (planned, #30) marker in an unrelated already-merged changelog fragment
• a gameable substring-only test replaced with regex assertions pinned to the actual prohibition sentences
• a code-fence guard test that only caught the literal ```bash spelling, generalized twice (once for other language tags, once for indented fences inside numbered-list continuation lines)
• this issue's own changelog fragment carrying the same imprecise wording the command file itself was already fixed to avoid
• a test that didn't actually pin the sentence its own fix commit added

Closes #30.

[AI-generated]