fix(AC-315): wrap all providers with RetryProvider for transient error handling

[jayscambler]

Summary

The RetryProvider existed with correct 500/429/timeout handling but was never wired in — create_provider() returned bare providers that crashed on transient errors.

Root Cause

L20 and L21 hit two consecutive Anthropic HTTP 500s and crashed immediately because:

# Before: bare provider, no retries
return AnthropicProvider(api_key=..., default_model_name=...)

Fix

All providers from create_provider() are now wrapped in RetryProvider:

# After: automatic retry with exponential backoff
return RetryProvider(AnthropicProvider(api_key=..., default_model_name=...))

Default: 3 retries, 1s initial delay, 2x backoff, 60s max delay.

RetryProvider._is_transient() checks for: 500, 502, 503, 504, rate_limit, 429, timeout, overloaded, capacity, connection, temporarily unavailable.

Providers wrapped

• Anthropic
• OpenAI-compatible
• Ollama
• vLLM

(MLX is local — no network retries needed)

Test plan

• 5 new tests: create_provider returns RetryProvider, retries on 500, gives up after max, get_provider wraps
• 3 existing tests updated for wrapped name format
• ruff clean, full suite 4400 passed

[linear]

AC-315 Bug: L20/L21 CLI runs fail on generation 3 with Anthropic HTTP 500 — need retry logic