Fix Grok MCP timeouts and ask compatibility

[RazonIn4K]

What changed

• enforce bounded per-attempt and end-to-end deadlines for chat, search, and search-enabled asks
• disable search retries by default and return explicit degraded fallback results within the request budget
• normalize reasoning_effort: "none" by omitting the unsupported xAI field
• sanitize Axios errors and bearer credentials before logging
• prefer local environment files before network-backed Doppler startup
• expose runtime timeout/retry status through grok_health
• document the timeout policy and add regression coverage

Why

The previous search path could wait through multiple 120-second attempts, causing MCP client timeouts. Separately, Grok 4.5 rejects the literal reasoning_effort value none, even though the MCP compatibility schema permits it.

Impact

Search and search-enabled asks now finish within a bounded server budget, provider failures are reported truthfully, startup is deterministic when local credentials exist, and grok_ask works for callers that select reasoning_effort: "none".

Validation

• npm run build
• npm test — 36 tests passed
• sh -n start-grok-mcp.sh
• forced provider timeout degraded safely in 37 ms
• live stdio MCP grok_ask returned GROK_ASK_FIXED
• restarted HTTP gateway grok_ask returned GROK_GATEWAY_FIXED