SPLAT JSON Schema + camelCase docs/examples + CI validation

[stackedsax]

What

The SPLAT JSON Schema, plus the docs/examples fix it surfaced, as one coherent
change. Supersedes #24 (its schema commit is included here).

1. Generate the schema from the Go types — internal/splat/schema reflects
   splat.Job into schema/splat.schema.json (draft 2020-12), with Quantity
   and Duration mapped to strings. Regenerate with make schema; a drift test
   keeps it in sync.
2. Fix the field casing — BAMMM parses SPLAT via json tags, so it only
   accepts camelCase, but SPEC.md and every examples/*.yaml used
   snake_case and were silently dropping those fields at parse time.
   Rewrote all first-class fields to camelCase (scheduler-specific
   extensions.* keys left as-is), and removed the aspirational
   placement.mpi_mapping from full-reference.yaml.
3. Gate it in CI — a Go test validates every example against the schema, so
   a stray or mis-cased field now fails go test ./... instead of silently
   losing data.

Why combined

The schema and the casing fix are two halves of one thing: the schema is only
correct/useful once the examples match it, and the CI gate ties them together.
Splitting them created a circular dependency (the gate needs both camelCase
examples and the schema), so they land together.

Testing

• go test ./internal/splat/schema — schema drift check, shape check, and all 9
  examples validate.
• Confirmed at runtime: converted examples now emit their fields (e.g.
  slurm-job.yaml → --cpus-per-task=4 --mem=8G), which they didn't before.
• Full suite / vet / gofmt clean.

New deps: github.com/invopop/jsonschema, github.com/santhosh-tekuri/jsonschema/v6.

🤖 Generated with Claude Code