refactor(llm): keep a single transport-level retry policy

[wolfy-j]

Cause

LLM retry had two implementations for one concern:

• Runtime layer: providers.open moved retry ({max_attempts, initial_delay}) out of the driver context into contract call options, so the runtime re-ran the whole driver call on a retryable error, including a streaming call that had already delivered chunks.
• Transport layer: the shared provider transport ({attempts, backoff_ms}), used only by the OpenAI and OpenAI-compatible clients. Claude, Bedrock and Google never received a retry policy, so a configured retry was silently ignored for them.

Change

• Only the transport layer remains. providers.open keeps retry in the driver context; llm.lua no longer copies provider retry into contract arguments (it reaches the driver through its context); a per-call retry is still hoisted and replaces the provider policy.
• transport.request_retry resolves the policy per request (request value replaces context, false sends once).
• Claude, Bedrock and Google clients and handlers apply it like OpenAI, OpenAI-compatible and TypeSafe. Retry happens before any response body is read, so streams are never replayed.
• Status handlers send a single probe.
• README documents the one retry option.
• Separate commit: the Bedrock embed handler read the hoisted timeout from the wrong table, so a per-call embed timeout never reached the client.

Tests

Converted: provider tests assert retry stays in the driver context and with_options is never used; the retry integration app asserts the driver receives retry through its context and runs exactly once per call. Added: client retry tests (503 then 200, no retry, per-request override) for Claude, Bedrock and Google; handler forwarding tests; single-probe status tests for all six drivers; Bedrock embed timeout test. Each failed before its implementation.

• make run-tests, make run-lint, make check-manifests: pass
• Live run with OpenAI, Anthropic and Gemini keys: 1410 passed, 0 failed