UnicDB v1.53.11

Robust text extraction from non-standard Lite Model responses

When the model produces 77 output tokens but UnicDB extracts 0 chars (as
happened with glm-5-turbo through some Lite Model proxies), the text was
missed because the response used a shape the parser didn't recognize.

Fix

• parseChatCompletionsResponse now handles message.content as either a
  string or an array of typed parts ([{type:"text", text:"..."}]).
• Both parsers fall back to extractAnyText, a recursive walker that
  concatenates any string in text-shaped keys (text, content,
  delta, output_text, message).
• The SSE aggregator's empty-text path now calls fallbackTextFromSseEvents
  before throwing, so unrecognized SSE shapes still surface their text.

Files

• src/ai/provider.ts — extractAnyText + fallbackTextFromSseEvents + parser fallback wiring