fix: shim reliability and protocol compliance overhaul

Addresses the most critical remaining issues in the provider shim layer,
building on top of #124 (recursive schema normalization + try/finally).

openaiShim.ts:
- Throw APIError via SDK factory instead of plain Error — enables retry
  on 429/503 (was completely broken: zero retries for all 3P providers)
- Guard stop_reason !== null before emitting usage-only message_delta
  (Azure/Groq send usage before finish_reason)
- Fix assistant content: join text parts instead of invalid as-string cast
  (Mistral rejects array content on assistant role)
- Expose real HTTP Response in withResponse() for header inspection
- Skip stream_options for local providers (Ollama < 0.5 compatibility)

codexShim.ts:
- Throw APIError at all 4 throw sites (HTTP + 3 streaming errors)
- Add tool_choice 'none' mapping (was silently ignored)
- Forward is_error flag with Error: prefix (matching openaiShim)

src/services/api/codexShim.ts

@@ -1,3 +1,4 @@
+import { APIError } from '@anthropic-ai/sdk'
 import type {
   ResolvedCodexCredentials,
   ResolvedProviderRequest,
@@ -234,7 +235,10 @@ export function convertAnthropicMessagesToResponsesInput(
           items.push({
             type: 'function_call_output',
             call_id: callId,
-            output: convertToolResultToText(toolResult.content),
+            output: (() => {
+              const out = convertToolResultToText(toolResult.content)
+              return toolResult.is_error ? `Error: ${out}` : out
+            })(),
           })
         }
 
@@ -453,6 +457,7 @@ function convertToolChoice(toolChoice: unknown): unknown {
   if (!choice?.type) return undefined
   if (choice.type === 'auto') return 'auto'
   if (choice.type === 'any') return 'required'
+  if (choice.type === 'none') return 'none'
   if (choice.type === 'tool' && choice.name) {
     return {
       type: 'function',
@@ -553,7 +558,13 @@ export async function performCodexRequest(options: {
 
   if (!response.ok) {
     const errorBody = await response.text().catch(() => 'unknown error')
-    throw new Error(`Codex API error ${response.status}: ${errorBody}`)
+    let errorResponse: object | undefined
+    try { errorResponse = JSON.parse(errorBody) } catch { /* raw text */ }
+    throw APIError.generate(
+      response.status, errorResponse,
+      `Codex API error ${response.status}: ${errorBody}`,
+      response.headers as unknown as Record<string, string>,
+    )
   }
 
   return response
@@ -633,11 +644,9 @@ export async function collectCodexCompletedResponse(
 
   for await (const event of readSseEvents(response)) {
     if (event.event === 'response.failed') {
-      throw new Error(
-        event.data?.response?.error?.message ??
-          event.data?.error?.message ??
-          'Codex response failed',
-      )
+      const msg = event.data?.response?.error?.message ??
+        event.data?.error?.message ?? 'Codex response failed'
+      throw APIError.generate(500, undefined, msg, {} as Record<string, string>)
     }
 
     if (
@@ -650,7 +659,10 @@ export async function collectCodexCompletedResponse(
   }
 
   if (!completedResponse) {
-    throw new Error('Codex response ended without a completed payload')
+    throw APIError.generate(
+      500, undefined, 'Codex response ended without a completed payload',
+      {} as Record<string, string>,
+    )
   }
 
   return completedResponse
@@ -806,11 +818,9 @@ export async function* codexStreamToAnthropic(
     }
 
     if (event.event === 'response.failed') {
-      throw new Error(
-        payload?.response?.error?.message ??
-          payload?.error?.message ??
-          'Codex response failed',
-      )
+      const msg = payload?.response?.error?.message ??
+        payload?.error?.message ?? 'Codex response failed'
+      throw APIError.generate(500, undefined, msg, {} as Record<string, string>)
     }
   }