Fix optional records and tuples in OpenAI and Anthropic structured output codecs - #10502

An optional property whose type is converted for the provider can't be decoded by the OpenAI or Anthropic structured-output codecs. This covers records, which become [key, value] pairs, and tuples, which become objects with numeric keys. The codec advertises the converted shape and then rejects that same shape.

Reproduce

import { Schema } from "effect"
import { toCodecOpenAI } from "effect/ai/OpenAiStructuredOutput"

const schema = Schema.Struct({
  value: Schema.optional(Schema.Record(Schema.String, Schema.Number))
})
const { codec } = toCodecOpenAI(schema)

Schema.decodeUnknownSync(codec)({ value: [{ 0: "a", 1: 1 }] })
// Throws: Expected array | null at ["value"]
// Expected: { value: { a: 1 } }

The same call succeeds when value is required. On 4.1.2 these also fail, with both toCodecOpenAI and toCodecAnthropic:

• optionalKey(Record)
• optional(NullOr(Record))
• an optional struct that contains a record
• optional(Tuple)

With @effect/ai-openai, a tool with any of these parameters fails the model's first call.

Fix

optionalToNullable reused the property's encoded form on the decoded side, so the conversion ran twice. It now uses SchemaAST.toType(type) for the decoded side. Types without a conversion are unaffected. New tests cover encoding and decoding an optional record and an optional tuple, including null and an absent key.

Validation

• pnpm test --run packages/effect/test/ai/OpenAiStructuredOutput.test.ts packages/effect/test/ai/AnthropicStructuredOutput.test.ts: 111 passed. Both new tests fail without the fix.
• pnpm lint and pnpm check pass.
• The repro above throws on 4.1.2 and passes with this change.

🤖 Generated with Claude Code

Closes EFF-3562