v0.3.0 - evaluated skills, tidied layout

All 34 skills now carry an evaluation manifest and every one of them has been
played through a real agent inside a throwaway fixture: 34/34 pass. The layout
was tidied at the same time, and the deprecated rules.ts is gone.

The most useful thing this release produced is a rule, learned the hard way
over six failing runs: a repo_clean allowlist only fits read-only skills.
A skill that builds something legitimately produces files nobody can enumerate
upfront (tests, configs, adapters, CI jobs, probe scripts), so the allowlist
never converges and each run punishes the agent for obeying its own skill.
Gate the invariant instead - what must not happen.

Added

• docs/developing.md: working on the toolkit itself - repo map, what each
  harness actually protects, the playground, and how to add every artifact
  type (hooks and agents had no guide at all).
• A header illustration on the README.

Changed

• The internal harnesses (skills.ts, tests.ts, mutations.ts, eval.ts)
  moved to tools/. install.ts, the only user-facing entry point, stays at
  the repo root.
• templates/ split by audience: it now holds only what a CONSUMER copies
  into their repo, while the artifact skeletons a CONTRIBUTOR starts from
  moved to docs/templates/.
• docs/writing-a-skill.md rewritten against the current contract. It
  described the pre-0.2.0 world, so a contributor following it produced a
  skill that CI rejects: no mention of the required eval.yaml, the per-skill
  README.md, the bilingual quoted triggers, or the test procedure that any
  script must ship.
• bug-triage-structured 1.1.0 and adversarial-feature-challenge 1.1.0:
  two protocol steps that evaluation runs showed agents skipping (the
  recurrence check, the persona switch) are now mechanical - a command to run
  and a count to state, rather than an instruction to remember.

Fixed

• templates/SKILL.template.md was still entirely in French and carried a
  frontmatter GitHub could not parse (an unquoted colon), so the repository
  home rendered a YAML error banner. Both are now guarded by
  tools/skills.ts templates, wired into npm test.
• Evaluation manifests no longer punish an agent for obeying its own skill:
  build output and dependency lock files are pruned from the snapshot, and
  repo_clean's except accepts a name-* prefix glob for skills whose
  deliverables cannot be named upfront (one ticket per finding).
• The writing-robust-tests fixture can actually run a test suite offline
  (node --test, nothing to install), so the "see it red, then green"
  discipline the skill demands is demonstrable rather than impossible.

Removed

• rules.ts, deprecated since 0.1.x and superseded by
  install.ts install <repo> --rules-only.