Fix cross-model auto title generation reliability

[ai-jiaqian]

What's fixed

• fix chat first-turn auto-title trigger reliability across conversation switches/resend/edit paths
• improve provider-agnostic title extraction for OpenAI/Anthropic/Google compatible responses
• add deterministic local fallback title generation in chat and agent flows
• add structured title diagnostics: http_non_200 / empty_content / parse_failed with status + dataPreview
• fix Anthropic-compatible gateway title request compatibility by removing thinking: { type: "disabled" } in title requests

Root cause

• title trigger could be skipped due to stale message state when switching conversations
• some gateways/models returned alternative payload shapes not covered by old parsing
• Anthropic-compatible gateway (qianfan coding endpoint) rejected title requests with 422 because thinking.type=disabled

Validation

• bun tests for title extraction/trigger/fallback/diagnostics all pass
• TypeScript checks pass for shared/core/electron

Notes

• openspec/ remains gitignored