fix(agent): send thinking:disabled for DeepSeek V4 and retry reasoning-only responses

[iojaffee]

What

Fixes #685 — the backend agent path never sent thinking: { type: "disabled" } for DeepSeek V4 models, so deepseek-v4-flash/deepseek-v4-pro could spend its entire max_tokens budget on reasoning_content and return an empty content, surfacing as LLM response did not contain assistant content in the UI chat.

Root cause

Three compounding issues in src-tauri/src/agent/provider.rs:

1. structured_task_config forced reasoning: auto, overriding the user's explicit reasoning choice from Settings. So even choosing "off" had no effect on backend agent turns.
2. apply_openai_reasoning returned early for custom providers, so the DeepSeek preset (provider: custom, api.deepseek.com) never received the thinking field the frontend sends (src/lib/llm-providers.ts:466-481).
3. No retry fallback: when the model produced only reasoning_content with no final content, both generate_openai_like and stream_openai_like surfaced the opaque error instead of retrying with thinking disabled (the frontend direct path already does this via isReasoningOnlyResponseError in chat-panel.tsx:1197-1208).

Changes

• apply_openai_reasoning: mirror the frontend DeepSeek handling — thinking: {type:"disabled"} on off; thinking: {type:"enabled"} + reasoning_effort on high/max; nothing on auto.
• structured_task_config: keep the user's explicit reasoning choice; fall back to the portable auto default only when unset.
• generate_openai_like / stream_openai_like: retry once with thinking disabled when the response contained only reasoning and no final content (guarded to DeepSeek V4 endpoints only).

Tests

Added 9 unit tests (all passing, 30/30 in agent::provider):

• deepseek_endpoint_detection, deepseek_v4_model_detection
• deepseek_reasoning_off_sends_thinking_disabled
• deepseek_reasoning_high_sends_thinking_enabled_with_effort
• deepseek_reasoning_auto_omits_controls
• structured_task_config_keeps_user_explicit_reasoning
• structured_task_config_defaults_to_auto_when_unset
• can_retry_thinking_only_for_deepseek_v4
• reasoning_only_error_detection
• Updated structured_task_config_raises_output_budget_and_keeps_explicit_reasoning (was: forced auto)

Verification

Reproduced against live api.deepseek.com before the fix: complex agent prompt + deepseek-v4-flash streaming → 6599 SSE chunks all reasoning_content, content stayed null, finish_reason: length. Same prompt with thinking: {type:"disabled"} → full content returned reliably. cargo check and all 30 provider tests pass.