Implement the Subject Sources foundation

[JeroenDeDauw]

Fixes #993

The ADR 23 foundation. With
only the local Source registered, every path resolves what it did before: nothing user-visible changes and nothing
stored changes.

What lands

T1 — SubjectId is a (source, localId) pair. A bare s + 14 nanoid characters still means a local Subject and
is still stored as that bare string; anything else is sourceKey:localId. Every place an id arrives as text — a REST
path, a revision slot, a relation target, a Lua call — goes through SubjectIdParser, which is handed the local Source
key and resolves nothing. Local slot JSON, Neo4j id and API output are unchanged, pinned by a byte-identical slot
test.

T2 — the Source contract, its registry, and LocalSource. One interface, one registry keyed by Source key, the
local revision slot refactored into the default registered Source under the MediaWiki Wiki ID. Extensions contribute
Sources through NeoWikiRegistrar::addSource(), alongside the Property Type and Page Property Provider registries.
ADR 23's "Source interface contract" open question is resolved in the ADR.

T3 — a Schema reference carries its Source. SchemaReference pairs a Source with the unchanged SchemaName, in
the two places a reference is persisted: a Subject's schema and a relation property's targetSchema. Resolution
routes through the Source the reference names, which need not be the Subject's.

T4 — a relation target naming an unregistered Source is an error-severity violation, raised on the write and
validate paths and never on read of persisted data. It blocks like every other error (ADR 26): under
$wgNeoWikiEnforceValidation, and only when the edit introduces it, so a Subject already carrying such a target stays
editable. Where one is stored anyway, it degrades on read — the projections drop the edge and the triple, and
resolution yields no Subject.

Design calls

• Serialized form sourceKey:localId, split at the first colon. Source key [A-Za-z][A-Za-z0-9_-]{0,63}; a
  foreign localId is RFC 3986 pchar, minus percent-encoding and minus & and ', capped at 256 characters. Every
  PHP grammar is \z-anchored, so a trailing newline cannot pass.
• The local key is the Wiki ID, and is exempt from that grammar. A wiki whose id is not a well-formed Source key
  still resolves its own bare ids; only the explicit <wikiId>:<localId> spelling is unusable there.
• An explicitly-local reference canonicalizes to bare, for Subject ids and Schema references alike, so one thing
  has one identity — which the id-keyed maps rely on.
• A foreign Schema reference persists as {source, name}. A bare string is always a local name, so a Schema title
  containing a colon (ISO:9001) stays valid.
• No client-side source validation (ADR 25). A server violation renders through the field-error path the editor
  already has.
• Sources register lazily, as closures built on first resolution: constructing the local Source reaches the
  subject-to-page index, which lives in the graph projection.
• isEditable() and foreign-source isValidLocalId() are contract surface with no consumer until sourced-subject
  rendering lands. They are frozen now because a Source author must answer them from the start.
• Foreign relation edges and unresolvable Schema references degrade — skipped with a warning — rather than
  project. The projection skip covers every foreign target, resolvable or not; T4 refuses only the unresolvable ones.
• RelationId shared SubjectId's $-anchor newline flaw, fixed here. Pre-existing, one line.

Considered, omitted

• A write-capability stub on Source. Write-back brings its own method when it is built.
• A TypeScript SubjectIdParser. Only canonical ids reach the frontend, so there is nothing to canonicalize there.
• Registry-aware colon splitting of Schema-name strings. Rejected: registering a Source later would retroactively
  reinterpret a Schema name that already contains a colon.
• Neo4j stub nodes for foreign relation targets. The stub would carry this wiki's wiki_id, and one without it scopes
  no query correctly.

Follow-ups filed

• Validate endpoint reveals existence and Schema of Subjects on unreadable pages #1266 — the validate endpoint is a read-authorization oracle
  (pre-existing).
• ViolationDiff treats a changed violation at the same index as not new #1267 — ViolationDiff identity ignores args.

Reviewing

The three commits stage the foundation — identity, the Source contract, Schema references and the target guard —
with the review findings from #1330 folded in and co-authored. PHP
and TypeScript assert the id grammars against one shared fixture, tests/vectors/subject-ids.json.

_{AI-authored — Claude Code, Opus 5 (max); detailed in-session spec from @JeroenDeDauw; diff not yet human-reviewed; phpcs, phpstan, PHPUnit and the TypeScript gates green locally at each revision, and all CI checks green on the current head.}

Production notes

Design and orchestration by Fable 5 (max); implementation and the post-review fix batch by Opus 5 (max) subagents.
Four independent AI review passes — code, security, tests, and a diff scan against a superseded branch — with every
blocking finding fixed before this description was written.

[JeroenDeDauw]

I did not fully review this yet

[JeroenDeDauw]

Note: before merging I'd like to ALSO finish my own review. Prior/additional review by others is useful and welcome

[JeroenDeDauw]

Thanks @alistair3149. I've resumed my review and identified some structural issues (not related to your changes). Will continue on this tomorrow.