Schematize `custom_metadata` and filter on its fields

[shangyian]

Summary

[Relates to: #1352]

Nodes carry a free-form custom_metadata JSON field with no schema enforcement and no way to query by its contents. We should add two capabilities:

• Schematize: let systems register a JSON Schema for an individual custom_metadata key, validated on write.
• Filter: let clients filter nodes by custom_metadata fields.

Design decisions

Per-key schemas, keyed (key, node_type?, namespace?): custom_metadata is a shared bag with many independent writers, so the unit of schema ownership must match the unit of write ownership: the top-level key. A single whole-object schema is a non-starter (no single
owner). Resolution picks the most specific active row per key: (key,node_type,namespace) > (key,·,namespace) > (key,node_type,·) > (key,·,·).

Lax, write-time-only validation: Keys with a registered schema are validated on write and unregistered keys are ignored. Registering a
schema never retroactively rejects or mutates stored rows. This keeps existing data working and requires no migration.

Two registration surfaces: Ad-hoc via REST API or durable via repo-backed deploy YAML (namespace-scoped), mirroring how nodes are authored. Globally-scoped keys are intended for cross-cutting/legacy keys but new integrations should claim a namespaced key.

Tiered filtering:

• [Tier 0] any key, no schema: equality/containment via a single global GIN index on the JSONB column (custom_metadata @> '{"k":"v"}'). Works for every key with no per-key setup.
• [Tier 1] schema registered: typed operators (numeric range/sort) backed by a per-key expression index, plus appearance in a curated, type-validated facet catalog.

Schemas are not a hard gate on basic filtering but they provide better guarantees around quality / discoverability.

Sub-issues

• Migrate custom_metadata to JSONB + global GIN index
• Add a per-key custom_metadata schema registry (table + resolution)
• Validate custom_metadata against registered schemas on write (lax)
• Filter nodes by custom_metadata fields (DB layer)
• Registry REST API: CRUD, facet catalog, violations, per-key indexes
• Expose customMetadataFilters on GraphQL findNodes
• Register custom_metadata schemas from repo-backed deploys
• Seed schemas for existing high-volume keys + document the model
• Follow-ups (pagination counts, dedup, index naming, CONCURRENTLY)