Bug

Graph entity extraction fails with 400 Bad Request when using OpenAI providers (gpt-4o-mini, gpt-4o).

Root Cause

GraphExtractor::extract() calls chat_typed_erased::<ExtractionResult> which uses cached_schema::<T>() to generate a JSON Schema via schemars::schema_for!(). The generated schema is sent as response_format: {type: "json_schema", strict: true, ...} to OpenAI.

OpenAI's structured output requirements for strict: true mode mandate:

1. additionalProperties: false on all object schemas — schemars does not add this
2. All fields must appear in required — optional fields (summary, temporal_hint: Option<String>) are generated as non-required by schemars

When these conditions are violated, OpenAI returns 400 Bad Request.

Reproduction

Run graph test with extract_model = "gpt-4o-mini" via ACP. Both conversation turns trigger extraction; all extraction calls fail with:

WARN zeph_memory::semantic: graph extraction failed: LLM error: OpenAI API request failed (status 400 Bad Request)

Graph works in CLI mode with Claude Haiku (Claude handles $defs/optional differently).

Fix

cached_schema::<T>() must post-process the generated schema before sending to OpenAI:

1. Recursively add "additionalProperties": false to all type: "object" schemas
2. Move all properties to required, and for Option<T> fields use anyOf: [{...T schema}, {"type": "null"}]

The Gemini provider already handles schema normalization via normalize_schema() in gemini.rs (PR #1635). A similar transform is needed for OpenAI's structured output path.

Alternatively: use response_format: {type: "json_object"} (no-strict mode) with explicit schema in system prompt for extraction — this avoids strict-mode requirements and works with all OpenAI models.

Impact

• All graph extraction via OpenAI providers is broken (silently fails with 400)
• Claude providers unaffected (structured output schema requirements differ)
• Discovered after fix(acp): apply graph memory config to ACP sessions #1645 fix made graph config reachable in ACP sessions

Severity

High — graph memory feature non-functional with OpenAI providers