Align AI schema $ref pointer escaping with updated STJ behavior

[Copilot]

System.Text.Json now percent-encodes JSON Pointer reference tokens when emitting $ref URI fragments. Our AI schema fix-up only applied ~0 / ~1 token escaping to the prepended parameter segment, which could produce under-encoded references inconsistent with the rest of the document.

• Schema $ref alignment

  • Update EscapeJsonPointerSegment to match STJ's current behavior:
    1. RFC 6901 token-escape (~ → ~0, / → ~1)
    2. URI-fragment percent-encode via Uri.EscapeDataString
  • Keep the existing $ref tail passthrough unchanged so exporter-produced inner pointer segments are not double-encoded.

• Regression coverage

  • Keep the existing a/b~c case to verify slash/tilde token escaping remains stable.
  • Add a percent-encoding case in both affected test suites using hello%20world, asserting the emitted segment becomes hello%2520world.

• Resulting behavior

  private static string EscapeJsonPointerSegment(string segment)
  {
      string escaped = segment.Replace("~", "~0").Replace("/", "~1");
      return Uri.EscapeDataString(escaped);
  }

  Example output change for a parameter named hello%20world:

  before: #/properties/hello%20world
  after:  #/properties/hello%2520world
  

[Copilot]

Pull request overview

Updates AI JSON schema $ref generation so JSON Pointer reference-token escaping matches System.Text.Json’s current behavior by applying RFC 6901 token escaping followed by URI-fragment percent-encoding. This prevents under-encoded $ref fragments when parameter names contain %-style sequences.

Changes:

• Updated EscapeJsonPointerSegment to token-escape (~//) and then percent-encode via Uri.EscapeDataString.
• Added regression coverage asserting % sequences are double-encoded in $ref fragments (e.g., hello%20world → hello%2520world) in both impacted test suites.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
test/Libraries/Microsoft.Extensions.AI.Tests/Functions/AIFunctionFactoryTest.cs	Adds regression test ensuring % in parameter names is percent-encoded in emitted $ref fragments.
test/Libraries/Microsoft.Extensions.AI.Abstractions.Tests/Utilities/AIJsonUtilitiesTests.cs	Adds regression test for percent-encoding behavior in AIJsonUtilities.CreateFunctionJsonSchema.
src/Libraries/Microsoft.Extensions.AI.Abstractions/Utilities/AIJsonUtilities.Schema.Create.cs	Updates JSON Pointer segment escaping to include URI percent-encoding after RFC 6901 token escaping.

[tarekgh]

The percent-encoding here matches the JsonSchemaExporter change in dotnet/runtime#130164, but that change is on runtime main (11.0) and this repo still pins System.Text.Json at 9.0.18 / 10.0.10 / 8.0.6, none of which percent-encode yet.

So until an STJ build with #130164 flows in, the prepended #/properties/{param} segment gets percent-encoded while the exporter-produced tail segments (nested $refs into $defs, etc.) stay un-encoded. For a ref like #/properties/{param}/{tail} that is inconsistent whenever the tail also contains characters that need encoding.

The top-level property segment is fine on its own since it points at a raw property key and percent-encoding is the correct way to reference it, so real impact is limited to exotic names. But should we hold this until the STJ bump lands, or is there a reason to take it ahead of the dependency? Just want to make sure we don't ship a half-encoded pointer in the interim.

[jozkee]

I thought it was fine to fix this ahead of time for MEAI, but I think the concern is valid. Let's wait until S.T.J ships #130164 and then conditionally align the fix with the STJ version — percent-encoding only when the runtime STJ also percent-encodes the exporter tail, so we don't ship a half-encoded pointer in the interim. Thanks.