The count_tokens handler was missing the model mapping step that the
messages and chat-completions handlers already had, so requests using a
mapped source name (e.g. claude-opus-4-8) failed model lookup and
returned a default token count, surfacing as "model not found" in
clients. Add the same mapping logic and register claude-opus-4-8.

Co-Authored-By: Claude <noreply@anthropic.com>

GitHub Copilot's claude-opus-4.8 rejects requests whose final upstream
message is not a user message ("assistant message prefill"). A trailing
Anthropic tool_result translates to an OpenAI `tool` role message, so
conversations ending in a tool call hit a 400. Append a placeholder user
message when the translated conversation does not end with a user turn.

Also switch local Claude CLI config to claude-opus-4-8.

Co-Authored-By: Claude <noreply@anthropic.com>

The assistant-prefill fix in f6afbdd used the condition `role !== "user"`,
which also matched tool-terminated turns. In the Claude Code agent loop, a
turn whose final user message carries only a tool_result translates to an
OpenAI `tool` role message, so every tool round hit the condition and got a
spurious `{role:"user", content:"Continue."}` appended — making the model
respond to "Continue." instead of the real tool result.

Verified directly against api.githubcopilot.com that claude-opus-4.8 returns
200 for tool-terminated conversations and only 400s on assistant-terminated
ones ("This model does not support assistant message prefill. The
conversation must end with a user message."). Narrow the condition to
`role === "assistant"` so only genuine prefill is patched.

Add regression tests covering both: tool_result-terminated must NOT append
Continue, assistant-terminated must.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>