docs: FAMILY.md v2 — member classes, and clause 7 stated per class

[chaoz23]

Resolves #80. This is a contract decision, so it is a PR rather than a push — the call is reviewable before it binds anything.

The problem

FAMILY.md v1 opened by saying it "describes the contract the family already implements." The audit found that false of clause 7: 2 of 4 D&D members ship no --pipe, and 1 ships no MCP server. The family's canonical contract overstated its own conformance — the same doc-truth defect these tools exist to catch, in the document that defines them.

The decision

The obvious fixes were both wrong. Marking clause 7 aspirational gives up on surfaces that genuinely should exist. Doing all the work means building --pipe and an MCP server for table-kit, which is transport plus a session ledger — --pipe is one-query-in / one-verdict-out, and that shape has no honest meaning for a thing that records a session. You'd get a surface built to satisfy a table.

So v2 states clause 7 per member class:

Surface	verdict	transport	adjacent
SKILL.md, tool.json, llms.txt, --schema flag	✅	✅	✅
--pipe	✅	—	—
MCP server	✅	—	—

• verdict — srdcheck, charactercheck, dmcheck. Every clause.
• transport — table-kit. Clauses 1 and 3 bind it where it emits a verdict (its own coverage/QC), not elsewhere.
• adjacent — inkcheck, loudcheck. Clauses 2–6, different domains, outside the D&D choreography and the honest lane.

That last class replaces v1's "recorded divergence" language. inkcheck and loudcheck were never breaking a rule — they were never in its scope, and describing them as an exception was the contract being imprecise about its own boundary.

What this does to the open issues

issue	disposition
chaoz23/dmcheck#14 — no --pipe	stays open — dmcheck is a verdict tool, the clause binds
chaoz23/table-kit#23 — --schema is a subcommand	stays open — binds every class, and it's a small alias
chaoz23/table-kit#21 — no --pipe	close as won't-fix if this merges
chaoz23/table-kit#22 — no MCP server	close as won't-fix if this merges

I have not closed anything. If you'd rather table-kit ship both surfaces anyway, reject this PR and all four stay as work.

Also corrected in v2

• Clause 1's honest lane was narrower than its own reference implementation. v1 said "outside the tool's codified jurisdiction"; srdcheck's refusal contract 1.1 also covers cannot answer from the evidence available (incomplete coverage), which is exactly what table-kit's qc/report exit-2 lane does too. Widened to match reality.
• Usage errors must not share the honest lane's exit code, stated explicitly, with srdcheck's exit 3 as the precedent. This is the rule behind [DMC-010] Exit 2 is overloaded between the honest lane and argparse usage errors dmcheck#15 and the HONEST_LANE_OVERLOAD check.
• Clause 7 gains a truth obligation for SKILL.md. Its acceptance test — "a fresh-context agent, given only this file, can produce a well-formed invocation" — checks invocability, and every member passed it while misstating its own exit contract. Claims must now be executable and executed, by scripts/family_conformance.py (feat: add a family conformance gate that executes documented exit contracts #82, merged).
• The two open clause-7 gaps are named inline, so the document can't drift back into claiming a conformance it doesn't have.

Per clause 7's own versioning rule, this is a heading bump (v1 → v2) with a changelog entry.

🤖 Generated with Claude Code