fix(sprout-agent): retry on pre-response transport errors

[tlongwell-block]

What

Broaden the LLM POST retry predicate in crates/sprout-agent/src/llm.rs from is_timeout() || is_connect() to also include reqwest's is_request(). Add a tracing::warn! on each retry so this isn't silent in production.

Why

The agent has been bubbling these to ACP instead of retrying them:

Agent reported error: {"code":-32000,"message":"llm: transport: error sending request for url (https://api.anthropic.com/v1/messages)"}

That exact "error sending request for url (...)" string is reqwest's Kind::Request Display. The current predicate misses it. In practice, is_request() covers the real-world transient failures we want to retry: TLS handshake aborts, sockets dropped or reset mid-send, h2 GOAWAY/RST_STREAM, hyper protocol errors. These are pre-response failures, and the LLM POST is idempotent — the body is serialized once before the loop and there's no observable side effect until a response is received — so replaying is safe.

Status-code retries (5xx, 429), auth bail-out (401/403), backoff, and MAX_RETRIES are unchanged.

Test

New post_retries_on_dropped_connection_before_response in llm.rs: a TCP listener that accepts and drops the first connection without writing any response bytes, then serves a normal JSON body on the second connection. Asserts the post succeeds and the server saw ≥2 connection attempts.

I verified the test reproduces the bug — with the old is_timeout() || is_connect() predicate it fails with literally:

Llm("transport: error sending request for url (http://127.0.0.1:xxxxx/v1/x)")

That's the same message Tyler reported from Anthropic.

Local: cargo test -p sprout-agent 17 integration + 44 lib tests pass; cargo clippy -p sprout-agent --all-targets -- -D warnings clean; cargo fmt -p sprout-agent --check clean.

Credit

Mari (GPT-5.5) led research and proposed the predicate. Sami (Claude Opus) verified all the load-bearing claims against the tree and sharpened the safety argument (body-serialize is outside the loop, so broadening to is_request() doesn't risk retrying a malformed body). I confirmed both, wrote the diff, and added the regression test.