T20: Design the target baseline schema

[thedancingdeveloper]

Not a translation. Because the project deploys fresh, the schema can be restructured for free — bake in the media-type-generic core (§5), the P5 profile-provenance tables, and the P6 explainable-decision records now. Doing this later costs a migration chain and a backfill.

Acceptance: schema reviewed and agreed before T21 lands.

[thedancingdeveloper]

This item is the gate the rest of P1 is waiting behind, and it needs a human

The design work is done; the acceptance criterion — "schema reviewed and
agreed" — is not, and cannot be done by an agent. Recording what a
reviewer would need, so it is actionable rather than just outstanding.

What exists to review, on feat/mariadb-baseline (PR #107):

• migrations/001_baseline.sql — 907 lines, 60 CREATE TABLE, replacing the
  18-file migration chain.
• docs/DATABASE.md — rewritten to describe it, so the schema can be read as
  prose rather than as DDL.

What it claims to embody, from §5 and §7 of docs/UNIFIED-ARR-PLAN.md:

• media_entities / media_files as the media-type-generic core;
• the five P5 profile/provenance tables;
• the two P6 decision-record tables.

Index parity was checked: 60 old named indexes against 55 new ones; all
eight missing names are covered by UNIQUE constraints, FK-implied indexes, or
leftmost prefixes.

What is blocked on this: T21, T22, T23, T24, T25, T26 (#59–#64). The work
queue carries depends_on: [T20] on all six, so nothing in the database swap
dispatches until this item is closed. The rest of P1 that does not touch the
schema — T27, T28 — is proceeding.

What "agreed" should mean here, given the plan's own framing that this is
free now and costs a migration chain plus a backfill later:

1. the media-type-generic core is the right shape for anime/music/books as
   first-class rather than forks (§5, P7);
2. the P5 and P6 tables are present and shaped so those phases do not need to
   alter them;
3. nothing that P3/P4 compatibility needs is missing.

A reviewer disagreeing on any of those is exactly why the gate exists — it is
cheaper to change 907 lines of DDL now than to migrate later.