Expose structured MCP reads with stable ledger identities

[Floating-Y]

MCP readers currently have to parse prose to recover facts and provenance.
Add structuredContent alongside existing text for find_entities,
search_chunks, entity_facts, and changes, using the same UUIDs as RDF exports.

• Preserve fact identities, both time axes, source document IDs, raw values,
  relation qualifiers, and explicit limit metadata.
• Return read failures as MCP errors instead of successful empty collections.
• Document the supported structured read contract and its limitations.
• Restore RDF links to derived premises in Turtle and JSON-LD.
• Add authenticated MCP integration tests and PostgreSQL CI coverage.

No database migration or new dependency is required.
Conflict/review export (#564), chunk provenance in RDF, and SPARQL
remain separate work.

Validation

• Formatting and workspace Clippy passed.
• Workspace tests with PostgreSQL required: 604 passed, 1 ignored.
• Frontend tests: 18 passed.
• Frontend production build passed.

Closes #550.

[WaylandYang]

Reviewed against dev after #600/#602 — no logical conflict with the qualifier work, identities are the real row ids, invalidated rows behave as the export does, and the tests read every id back. Thank you; merging.

Three small things I'll take as a follow-up rather than hold this on (tracked in an issue):

1. tools_graph.rs resolve() maps a database error to "could not look up entities", which entity_facts then wraps as Invalid entity: … (expected a name, or the uuid returned by find_entities) and neighbors / timeline / paths_between report as Unknown entity with isError:false. An outage reads as a bad argument; a distinct "could not read" result, as find_entities already returns, is the fix.
2. docs/decisions/0020 gets an inserted paragraph without a date; the records README asks for a dated revision note in place.
3. web/src/docs/mcp.md tabulates every structuredContent field under a heading called "contract" but never says whether a removal or rename needs a record. One sentence either way.

Also noted: refused calls (the shared check_call guard) skip the mcp.tool_called audit row; search_chunks.limit_reached and an entity-valued qualifier are untested.