fix(schema): strip null from enum arrays in tool schemas for strict validators

[Crazy-LittleBoy]

Problem

When using Kimi (Moonshot) models via Anthropic-compatible or OpenAI-compatible API endpoints, requests fail with:

400 Bad Request: parameters is not a valid moonshot flavored json schema,
details: <At path 'properties.output_mode.enum': enum value (null)
does not match any type in [string]>

Root Cause

schemars' AddNullable transform adds "nullable": true and removes "null" from the type keyword, but does not remove null values from enum arrays. This produces invalid JSON Schema where enum contains null while type is e.g. "string" — a contradiction that strict validators (Moonshot/Kimi) reject.

The existing cleanup in enforce_strict_schema only ran in strict mode, which is gated to specific provider IDs (FIREWORKS_AI, XAI, kimi_coding, etc.). Generic OpenAI-compatible and Anthropic-compatible endpoints never triggered the cleanup.

Fix (3 layers, architectural root cause)

1. Domain Layer — NormalizeNullable transform

• New transform in forge_domain/src/tools/definition/tool_definition.rs
• Replaces AddNullable in ToolCatalog::schema()
• Delegates type normalization to AddNullable, then recursively strips null from enum arrays when nullable: true is set
• The nullable marker alone fully conveys nullability (OpenAPI 3.0 convention); null in enum is redundant

2. App Layer — enforce_strict_schema defense-in-depth

• Null-stripping from enum moved outside the strict_mode gate in forge_app/src/utils.rs
• Guards external MCP tool schemas that bypass NormalizeNullable
• In strict mode: null stripped first, then anyOf conversion (unchanged behavior)
• In non-strict mode: null stripped, nullable marker preserved

3. Anthropic Pipeline — tool schema normalization gap fill

• EnforceStrictObjectSchema in forge_app/src/dto/anthropic/transforms/enforce_schema.rs now normalizes tools[*].input_schema
• Previously only processed output_format schema, completely skipping tool schemas

Testing

• forge_domain: 611 tests pass (3 new tests for NormalizeNullable)
• forge_app: 723 tests pass (2 new tests for non-strict null stripping + Anthropic tool normalization)
• Snapshot updated: null removed from output_mode enum in tool definition snapshot
• All existing strict-mode tests (Fireworks, OpenCode Zen, xAI, kimi_coding) pass unchanged

Co-Authored-By: ForgeCode noreply@forgecode.dev

[CLAassistant]

All committers have signed the CLA.