Keep tool argument headers and bodies in sync

[quinnj]

Modern call_tool encoded known tools' arguments twice: once for Mcp-Param-* headers and again for the request body. Custom JSON lowering could therefore send a routing header such as west with body value east, causing HTTP 400 / HeaderMismatch (-32020) before the tool ran.

Prepare one validated argument snapshot and reuse it for both representations through the existing JSON.JSONText API. Full object validation, duplicate-key/null semantics, schema discovery, omitted arguments, legacy calls and generic JSON-RPC behavior stay covered. This repairs the existing 2026-07-28 header/body contract; no exports or dependencies are added.

Validation:

• Full suites: 607 checks on Julia 1.13.0; 601 on Julia 1.10.12.
• 125 focused wire/error/side-effect checks also pass with HTTP 1.11.0 and JSON 1.0.0. The changing-value case fails against main with HTTP 400/-32020.
• Official conformance runner 0.2.0-alpha.11: custom headers 18/18 and invalid tool headers 11/11.
• Existing static-server JuliaC --trim=safe compile/run and Documenter build/doctests pass. The native gate does not cover the dynamic modern client.
• Hosted CI: all 10 jobs and the documentation status are green at 9b31da6f65ba2510ce90a576c20f66748f7c8604, including Linux/macOS/Windows on minimum/current/nightly Julia.

Alternating same-process loopback calls, cached schema, HTTP 2.8 / JSON 1.9, median of 31 pairs:

Julia	Routed arguments	Full call	Allocated bytes
1.13	10,000 integers	2.126 → 1.928 ms	1,319,616 → 1,326,784
1.10	10,000 integers	2.418 → 2.195 ms	4,592,176 → 2,837,408
1.13	1 MiB string	3.474 → 3.284 ms	essentially unchanged
1.10	1 MiB string	4.091 → 3.885 ms	essentially unchanged

These measurements include client/server JSON handling and actual TCP, excluding discovery, TLS and application work. Small warmed calls were unchanged. The first routed call added approximately 9 ms on Julia 1.13 and 22 ms on 1.10 in three fresh-process pairs; current-runtime numeric calls allocate about 7 KB more. Unrelated host work remained active, so these are bounded measurements, not a general throughput claim.

Co-authored by Codex