Marks the family-reference agent sections

[johnnyt]

Why

This repo is the de-facto reference for the family's agent-facing conventions,
but that status was implicit and the copies drift. Beads in the five sibling
repos currently say some version of "copy statifier-ex's, adapted down", which
is only safe if what to copy - and what not to - is written down here rather
than inferred by whoever reads the file next.

What

Adds docs/family-reference.md and points CLAUDE.md at it from two places:
the paragraph that already describes the .claude/wurk/ extension contract,
and the read-before-deciding list.

The doc marks each agent-facing section copy-verbatim, adapt-per-repo,
or not reference, and says what an adopter must change:

• Agent authority - adapt-per-repo, and the worked example of why the
  distinction matters. This repo's grant is standing; statifier_persistence,
  statifier_oban and opentelemetry_statifier hold a consent-scoped grant.
  Copying verbatim in either direction would silently change a repo's
  authority level. What is common is the structure: the per-action table, the
  anti-inference paragraph, the two override paragraphs, the reversibility
  principle, and the closing widening line.
• Beads that span repositories - adapt-per-repo.
• Non-interactive shell commands - copy-verbatim.
• The ExQuality block and the deps/ex_quality/usage-rules.md pointer -
  copy-verbatim; statifier-ui's copy is byte-identical today and that is the
  correct state.
• The .claude/wurk/ extension mechanism - adapt-per-repo. The contract
  is the reference; the per-repo count of override files is not, and should
  not be levelled up to match.
• The custom gate stages - not reference. The gate guard, the ADR guard,
  the ADR judge, the regression ratchet, and mix gate.check with its guard
  ledger stay out of scope for copying, so the scaffolds' "deliberately
  smaller than statifier-ex's gate" comments stand unreversed.
• mix gate.verify - the one exception, adopted by px-591, sui-4py,
  sp-7cu, sob-ehl and ots-4l6. It adds no stage to .quality.exs, so the
  smaller-gate statement stays true; it reads nothing but ex_quality's own
  report; and without it the wurk pipeline reports attested: false and the
  commit skill refuses, which is a broken unattended commit path. The doc
  records that it is being extracted upstream into ex_quality (st-wgr0), so
  adopters should expect to collapse their copies later.

Notes

• Docs only. gate.rb reports the carve-out (applicable: false, nothing
  under lib/, test/, config/, mix.exs, mix.lock), but
  mix quality --profile merge was run unconditionally per
  .claude/wurk/mr.md and is green across every stage. The ADR judge skipped
  cleanly - no file in this diff is in a judged ADR scope.
• No changelog fragment: agent-facing documentation is invisible to anyone
  using the public API.
• The consent-scoped exemplar is cited as statifier_persistence's
  CLAUDE.md on main plus its request numbers, deliberately not as a commit
  SHA - two SHAs written down earlier in this standardization pass were
  orphaned by non-fast-forward merges.
• Nothing in the doc changes this repo's authority table or its gate.

Closes st-i0qa