Summary

Closes #2007

Custom endpoint users can now configure context length during setup wizard instead of relying on slow probe-and-stepdown at runtime.

• Auto-detects from endpoint /models metadata when available
• Falls back to manual prompt when endpoint doesn't report it (with note that auto-detect at runtime may be slow)
• Saves value to context_length_cache.yaml with normalized base URL (rstrip("/")) for reliable cache hits at runtime
• Applies consistent range (1,024–10,000,000) and type (type(x) is int) guards for both auto-detected and manual values

Test plan

• test_custom_endpoint_auto_detects_context_length — endpoint reports context_length → saved automatically
• test_custom_endpoint_manual_context_length_prompt — no metadata → user enters "32768" → saved
• test_custom_endpoint_context_length_blank_skips — user leaves blank → skipped
• test_custom_endpoint_context_length_invalid_input — user enters "abc" → warning, skipped
• test_custom_endpoint_trailing_slash_normalized_in_cache_key — /v1/ → /v1 in cache key
• test_custom_endpoint_auto_detect_out_of_range_skips — 99M tokens from endpoint → warning, skipped
• All 16 targeted tests pass (test_setup + test_setup_model_provider), 0 regressions

CI note

CI failures are pre-existing on main (tests/tools/test_delegate.py — _saved_tool_names NameError, 20 tests). Unrelated to this PR.