docs(prd): PRD-0001 — citation verification for revision review

[tieguy]

The first PRD under the process proposed in #2 (this branch is based on it, so
the diff here is just the one new doc).

Proposes an operator-facing capability: when reviewing a revision that adds or
changes a citation, verify whether the cited source actually supports the claim
and show a categorical verdict — read-only, never an autonomous write, and it
abstains when it can't fetch or read the source.

It exercises the whole process end to end:

• a testable definition of done (incl. the load-bearing rule: never supported
  without a verbatim, in-session-locatable passage);
• four spawned ADRs it would trigger (verdict/action semantics, verification
  contract, crate boundary, source-snapshot storage);
• open questions that each carry a strawman answer to react to.

For discussion — not a request to build yet.

Validation: docs only, no code paths touched; no checks run.

🤖 Generated with Claude Code

[schiste]

How much details do we want in PRDs? Quick though on that one (and perhaps it should be an ADR), I think the tool needs to be as auditable and determinist as possible.

So for a citation thing like that, I would tend to create an abstraction layer where SP42 generates a JSON payload, model enrich/use it and returns a JSON (I would store them for research purposes but that could be an optin feature to make it easier).

But maybe this shouldn't be specific to this PRD ?

[tieguy]

I'm not sure. My gut is "PRDs are mostly for documenting user impact, ADRs are for technical details" but this is experimental for me too :)