…hat client

The setup wizard accepts whatever base URL the user pastes and probes it
with GET /models. The conversation-time client, however, concatenated
"" + "/v1/messages" for Anthropic and "/chat/completions" for OpenAI
against that same stored URL. For Anthropic providers the two paths could
diverge:

  - User pastes "https://proxy.example.com"        (root form):
    probe hits https://proxy.example.com/v1/models (no /v1 in base)
    chat client hits https://proxy.example.com/v1/messages  (root + /v1)
    — aligned.

  - User pastes "https://proxy.example.com/v1"     (OpenAI form):
    probe hits https://proxy.example.com/v1/v1/models (404)
    chat client hits https://proxy.example.com/v1/v1/messages (404)
    — silently broken.

The fix has two parts:

  1. Anthropic client (internal/provider/anthropic/anthropic.go):
     New() now strips a trailing /v1 from the stored base URL before
     storing the root, so +/v1/messages is correct for both shapes.

  2. Wizard probe (internal/cli/cli.go):
     Two openai.FetchModels(ctx, baseURL+"/models", ...) call sites in
     the custom and Anthropic wizard flows are replaced with a new
     fetchModelListCompat helper that walks the full set of candidate
     URLs from config.BuildModelFetchURLs (root, /v1, known compat
     suffixes). Endpoint-miss on one URL falls through to the next;
     non-miss errors bubble up immediately.

Tests:
  - TestBaseURLNormalizedForV1Messages (provider/anthropic): 6 table
    cases covering plain root, /v1, /v1/, /api/v1, trailing slash only,
    and empty (falls back to default).
  - TestFetchModelListCompatWalksCandidates (cli): 4 subtests using
    httptest — Anthropic root form falls through to /v1/models,
    versioned v1 base URL hits /v1/models directly, all-miss returns
    empty for manual fallback, connection refused bubbles up.