SchemaSerializer.Load throws ArgumentException on a .schema.json with an absolute outputPath or data source file, crashing kschema and the editor's File > Open #299

What's wrong

outputPath and file are deserialized through the semantic path types. An absolute value such as "outputPath": "/tmp/out" or "file": "C:\\data\\items.json" makes the converter throw ArgumentException.

SchemaSerializer.Load catches only JsonException (Schema/SchemaSerializer.cs:157 and :205). Open PR #297 adds NotSupportedException, but not ArgumentException. So the exception escapes to two callers:

• kschema validate / kschema generate: neither SchemaCommandLine.TryLoad (Schema/Cli/SchemaCommandLine.cs:193) nor Schema.Tool/Program.cs catches it. The user gets a stack trace instead of "Could not open ...".
• The editor: SchemaFile.Load (Schema.Editor/SchemaFile.cs:62-80) catches only IOException and UnauthorizedAccessException. Opening such a file, including from the recent-files list, throws out of the frame.

docs/schema-format.md says these paths are relative. Still, a hand-written or tool-written file with an absolute path is easy to produce. It should be refused as InvalidJson, not crash the process.

Repro

SchemaSerializer.Load(
    "{\"formatVersion\":1,\"codeGenerators\":[{\"name\":\"G\",\"language\":\"CSharp\",\"outputPath\":\"/tmp/out\"}]}",
    "/tmp/x.schema.json".As<AbsoluteFilePath>());
// throws ArgumentException: Cannot convert "/tmp/out" to RelativeDirectoryPath

Suggested fix / acceptance criteria

• Load returns SchemaLoadStatus.InvalidJson with a message that names the property and says paths must be relative to the schema file. Either:
  • catch ArgumentException around the deserialize call, or
  • give the path properties a converter that reports a JsonException with the JSON path.
• Tests for an absolute outputPath and for an absolute data source file:
  • Load returns InvalidJson;
  • kschema validate exits 1 with "Could not open".
• Optional: state in schema-format.md that absolute paths are refused.

This is related to #242 and #287, which are other load crashes on different inputs. PR #297 does not cover this one.