fix(relay): detect CRLFCRLF across chunks + cap drain buffer

[intendednull]

Why

bootstrap drain loop in handle_bootstrap_request_after_line reads 1 KB chunks looking for HTTP end-of-headers \r\n\r\n. but chunk.windows(4).any(...) only scans latest chunk. marker straddling two reads = miss → loop spins until BOOTSTRAP_IO_TIMEOUT (5s). also no cap on chunk count: peer can stream forever w/ no marker.

low severity (path is rate-limited by MAX_CONCURRENT_BOOTSTRAP_CONNECTIONS semaphore + 5s deadline), but a robustness gap worth closing.

Fix

accumulate every read into bounded Vec + search whole buffer each iteration → cross-chunk marker found. cap = BOOTSTRAP_DRAIN_BUFFER_CAP = 8 KiB, matches PROXY_REQUEST_LINE_BUFFER so relay applies one symmetric header budget. exceed cap → tracing::warn one line + Err(InvalidData) → dispatch_connection closes connection. seed accumulator w/ trailing 3 bytes of already_read so marker straddling request-line/chunk boundary is also caught.

rejected: rolling 4-byte tail across iterations. simpler in bytes but harder to reason about (short reads, seeded state, off-by-one risk). 8 KiB cost trivial vs clarity gain.

Verify

• cargo fmt --check clean
• cargo clippy -p willow-relay --all-targets -- -D warnings clean
• cargo test -p willow-relay 12/12 pass (10 existing + 2 new)
• new test proxy_bootstrap_detects_crlf_crlf_split_across_chunks — sends \r\n\r then pause then \n; asserts response in <2s (would hang to 5s deadline w/o fix)
• new test proxy_bootstrap_closes_when_drain_buffer_cap_exceeded — streams 16 KiB no marker; asserts EOF in <2s w/ no 200 OK

Closes #238

---

Generated by Claude Code