The previous README predated a cross-vendor standard and stated that none
existed. The Agent Skills specification (https://agentskills.io/specification)
now fills that gap, so this rewrites the README around it.

Adds:
- what a skill is, and the progressive-disclosure model agents use
- directory layout (SKILL.md, scripts/, references/, assets/)
- the full frontmatter table with the specification's constraints, including
  the name regex and the rule that name must match the directory
- guidance on writing the body: explicit decision rules, stop conditions, and
  restating the trigger outside the frontmatter
- size budgets per load stage, and the 500-line SKILL.md ceiling
- validation via skills-ref
- ECMWF house rules layered on the specification: report-and-gate,
  model-agnostic, never invent or hide findings, redact secrets, British
  English, project-agnostic
- installation paths per harness, preferring the vendor-neutral
  .agents/skills/ location, plus a table of per-product setup documentation
  from https://agentskills.io/clients
- a checklist for contributing a new skill

Ports five workflows into project- and language-agnostic skills written to
the Agent Skills specification. Each discovers the repository's own build,
test and documentation layout rather than hard-coding paths, tools or
organisation names.

- review-pull-request — works through review feedback until it converges.
  Merges two prior workflows (respond-to-review and the Copilot review loop),
  since the mechanics are identical and only the reviewer differs. Records the
  hard-won bot mechanics: the literal `[bot]` reviewer suffix is the only
  request form that works, and the `requested_reviewers` list cycling back to
  empty is the terminating signal. Notes that `gh pr view` does not show inline
  comments, so the GraphQL reviewThreads query is required. Never merges.
- doc-fact-check — executes runnable doc examples and verifies falsifiable
  claims against source. Keeps the distinction between a documentation bug and
  an environment problem, and refuses to invent imports to make a block run.
  Proposes fixes without applying them.
- onboard-repository — read-only survey producing a structured mental model.
  Trusts code over documents and reports the drift between them.
- robustness-audit — merges error-handling and edge-case auditing, which belong
  together: an unhandled edge case usually surfaces as a bad error. Severity
  ladder led by silent-wrong-answer. Asks before changing ambiguous or public
  behaviour.
- improve-code-coverage — treats coverage as a proxy rather than a goal.
  Categorises each gap (testable / defensive / dead / environment-specific),
  prioritises by risk over line count, and deletes dead code rather than
  suppressing it.

Also splits the README index into governance skills (which gate and never act)
and development skills (which may change code, but still never merge, publish
or release, and ask before changing public behaviour).

All five validate against the specification: name matches directory, name
regex and length, description present and within 1024 characters, no unknown
frontmatter keys, and each SKILL.md under the 500-line ceiling.