#348 Scaffold new extensions to publish-ready quality (12/12) by default

Problem

Following the documented swamp-extension-model workflow today produces an extension that scores 4/12 (33%) on swamp extension quality. The missing factors are not bugs or stylistic preferences — they're explicitly client-earnable items in the rubric:

• has-readme (2 pt) — no README written
• readme-example (1 pt) — README has no fenced code blocks
• rich-readme (1 pt) — README is short
• symbols-docs (1 pt) — no JSDoc on exported symbols
• has-license (1 pt) — no LICENSE file
• repository-verified (2 pt) — repository: field unset in manifest
• (often) fast-check (1 pt) — re-exporting bare z.enum(...) / z.object(...) fails deno doc --lint

This is a real friction point. After creating @magistr/telegram-send end-to-end today, getting from 33% → 100% took an extra round trip: write README, write LICENSE, add module + symbol JSDoc, restructure exports to satisfy deno doc --lint, add a verified repository URL, re-format, re-score, dry-run, push.

All of that work is mechanical and predictable — it could be done by scaffolding at extension-creation time.

Proposed solution

Add a swamp extension new <name> (or init) subcommand that scaffolds a publish-ready skeleton from a single prompt:

swamp extension new @magistr/foo
# prompts: kind (model/workflow/vault/driver/datastore/report), license (MIT default),
#         repository URL (optional, validated against the verified-host list)

The resulting directory should reach 12/12 on swamp extension quality immediately, with TODO: markers in the spots that require human input (description, code example, repository URL if skipped).

Concretely, the scaffold should produce:

1. manifest.yaml with name, version (from swamp extension version), description: TODO, repository: (if supplied), additionalFiles: [README.md, LICENSE.md], and the right typed-content array seeded with one entry.
2. README.md skeleton that already meets rich-readme and readme-example thresholds — sections for Install, Setup, Usage with two fenced code blocks of placeholder commands, ≥500 characters of structure.
3. LICENSE.md populated from the chosen SPDX identifier (MIT default), with current year and the authenticated collective name auto-filled.
4. .gitignore and deno.json matching the existing per-extension layout convention (good-planning / issue-lifecycle style).
5. For a model: extensions/models/<snake>.ts whose export const model carries a module-level JSDoc with @example, and whose any-exported Zod schemas have explicit type annotations so deno doc --lint passes (fast-check). The current swamp-extension-model Quick Start exports model only — that's fine; but if it grows to export schemas they should be annotated, and the template should demonstrate the safe pattern (or simply keep schemas un-exported).

Alternatives considered

• Document-only fix. Update the swamp-extension-model SKILL.md Quick Start so it ships a polished file by default (README + LICENSE + JSDoc + repository). This raises the floor without new CLI surface area, but Claude/users still have to copy and customize four files manually for each new extension. A scaffolder is strictly better, but the SKILL.md update is a worthwhile first step on its own.
• Quality-as-warning during extension push. Today quality is opt-in and push doesn't surface the score. Making push print a single-line "X/12 (Y%)" alongside the dry-run summary would at least make the gap visible at the moment authors care most. (Suggesting this in addition to scaffolding, not instead.)
• Defer to per-collective templates. Some collectives might want their own LICENSE / README header. The scaffolder can accept --template <path> later; for v1 the built-in MIT/skeleton is enough.

Additional context

• Existing swamp-extension-publish skill already encodes the publishing state machine and surfaces quality as optional. A new command would close the loop on the other end — authoring — so authors don't only learn about the rubric at push time.
• The repository-verified factor specifically requires a public HTTPS URL on github.com/gitlab.com/codeberg.org/bitbucket.org. The scaffolder should validate the host up-front rather than letting the failure surface at push.
• Real example for context: @magistr/telegram-send v2026.05.13.1 (https://swamp-club.com/extensions/@magistr/telegram-send) — went from 33% → 100% by adding the items listed above, no code or behavior change.