[codex] Normalize MCP schema integer constraints

[lawyer112]

Summary

• canonicalize integer-valued JSON Schema constraints emitted by dynamic MCP tools before provider dispatch
• recurse only through JSON Schema child-schema positions so examples, defaults, and unrelated numeric keywords are not rewritten
• add a regression test for Cursor app-control's move_agent_to_cloned_root schema and a preservation test for non-integer values

Problem

Cursor app-control can send MCP input schemas through a protobuf Struct. Protobuf numbers are decoded as floating-point values, so integer-only JSON Schema keywords such as minItems: 1 can be serialized as minItems: 1.0.

Strict OpenAI Responses-compatible validators reject the entire tool definition before the model can answer. The observed failure was:

invalid constraint value at #/anyOf/1/properties/rootPaths: minItems must be an integer, got 1.0

The existing MCP normalizer already repairs the missing root type: object annotation used by this app-control schema, but returned before normalizing constraints on otherwise valid object schemas.

Fix

Normalize non-negative, integral floating-point values for JSON Schema's integer-only size/count keywords into integer JSON numbers. The traversal is schema-aware and covers nested schemas in properties, unions, definitions, array items, and related schema containers.

Fractional values and other numeric constraints such as minimum: 1.0 remain unchanged.

Validation

• cargo test --package cursor-server --lib — 137 passed
• cargo clippy --package cursor-server --all-targets -- -D warnings
• cargo fmt --all -- --check
• git diff --check

The full cargo test --package cursor-server run also passed all suites except every_captured_mode_owns_and_renders_its_runtime_template. That assertion expects LF-only prompt text and fails on this Windows checkout; the identical failure was reproduced in a separate, unmodified origin/main worktree.