Reject incomplete structured LLM streams

[RANGROO]

Problem

OpenRouter could end a streamed structured response after only a JSON prefix such as { "character":. StobeServer treated any non-empty HTTP 200 stream as successful, then fell back to speaking and persisting the raw fragment.

Root cause

• the OpenAI-compatible streaming connector discarded provider error, finish_reason, and native_finish_reason metadata
• the structured dialogue path accepted a JSON-looking fallback even when it had no usable message

Changes

• capture and audit provider stream errors and finish reasons
• reject normalized truncation, content-filter, safety, and error completions
• reject JSON-looking structured responses without usable dialogue
• preserve plain-text provider fallback and salvageable partial JSON that already contains dialogue
• extend the existing structured-dialogue regression for the reported fragment

Compatibility

No database migration or configuration change. Existing plain-text fallback and usable partial structured responses remain supported.

Validation

• PHP lint: connector/openaijson.php, lib/chat_helper_functions.php, tests/structured_dialogue_contract_regression.php
• focused stream-failure classification probe passed, including length and Gemini-style MAX_TOKENS
• STOBE_DB_NAME=stobe_test php tests/structured_dialogue_contract_regression.php
• STOBE_DB_NAME=stobe_test php tests/rechat_flow_regression.php
• STOBE_DB_NAME=stobe_test php tests/chat_flow_regression.php
• git diff --check
• outgoing diff secret/local-path scan

Deployment

Not deployed.

Untested limits

No live OpenRouter request or in-game Kenshi retest was performed. If a provider fails after complete dialogue sentences have already streamed, those earlier sentences cannot be retracted; the failed completion is still rejected and no raw JSON fallback is accepted.