fix: pool Hive API connections in a shared keep-alive session

[danceratopz]

Replace the bare requests.post/requests.delete calls in HiveTestSuite.start/end, HiveTest.start/end and HiveTest.register_multi_test_client with a single process-wide requests.Session mounting an HTTPAdapter with a urllib3.Retry policy (connect-error retries with exponential backoff).

Why

See recent sporadic fails on the enginex dashboard:

• https://hive.ethpandaops.io/#/group/generic?suites=eels%2Fconsume-enginex%2Ceels%2Fconsume-engine
  
• https://hive.ethpandaops.io/#/compare/generic?runs=1786084315-24e13e88989d0eaef9e24bf9e63f561f,1786001568-2d5402ea56c410854dd4b539aeb83729,1785913833-8eb28d029fb94b74bc592dbd2f2d65e3,1785828828-a24ae793c6064646bb90caa87e5266d5,1785744108-f602da91d75d9e8b90e3e03d15550287,1785655766-004419ff491817b3ec9c6a3207a12425,1785569573-c9e187a1160aa8692a2386337bb47368,1785483617-93c7749ccd21ffaa669a80f3cc8f76d0,1785395624-30548b6eabaa72a83f4eff5074825b6a,1785309844-922f77c7c9f8cdd0bad3e08b2171d9c6

Each bare call opens a fresh TCP connection and leaves the socket in TIME_WAIT for 60 s. Ephemeral-port exhaustion is per destination tuple, and every Hive API call targets the one fixed endpoint, so all churn piles onto a single (dst_ip, dst_port). At consume-enginex throughput (~110–145 tests/s on the dashboard host × 3 calls/test ≈ 330–440 connections/s) the ~28k-port ephemeral range toward the Hive API fills in ~90 s and connect() intermittently fails with OSError: [Errno 99] (EADDRNOTAVAIL) on all xdist workers at once. In the 2026-07-18/19 nethermind eels/consume-enginex dashboard runs this produced 4–5 spurious "Test was terminated by host" failures and silently dropped 85/40 test results.

Pooling collapses ~3 connections/test (~163k per full run) into a few persistent keep-alive sockets — removing the mechanism rather than raising the ceiling.

Retry semantics

Only connect errors are retried (5 attempts, 0.5 s backoff factor, ~15 s total): they are raised before the request is sent, so a retry cannot duplicate a non-idempotent POST (no double start_test). Read errors and HTTP error statuses are not retried.

Keep-alive caveat (verified)

Pooling only helps if the server honors keep-alive. Hive's Go backend serves the sim API with a stock net/http server (internal/libhive/run.go: &http.Server{Handler: tm.API()}) — keep-alives on by default, no idle timeout, no handler emits Connection: close. Confirmed empirically below.

Validation

Hive --dev + nethermind 1.40.0-unstable, 12,888 consume-enginex Cancun tests at -n 4, sampling ss -tan state time-wait '( dport = :3000 )':

Metric	0.1.0a5 (per-call)	this PR (pooled)
Verdicts	12,888 passed	12,888 passed
Peak TIME_WAIT → hive API	9,684	66
Established → hive API	≤4 transient	8–9 persistent
Wall clock	36:04	36:58 (noise; perf ≈ neutral)

Notes

• One session per process = one per pytest-xdist worker; urllib3's pool is thread-safe.
• HiveTestSuite/HiveTest dataclasses unchanged (still asdict/JSON round-trippable, as execution-specs relies on).
• urllib3>=1.26.0,<3 declared explicitly (>=1.26.0 for the modern Retry kwargs; <3 congruent with requests>=2.31.0, and the 2.x line is what CI and execution-specs actually resolve).
• New unit tests (src/hive/tests/test_session.py) run against a local in-process HTTP server — no hive required: singleton behavior, retry config, 42-calls-over-1-connection reuse, connect-retry until a late-starting server is up.

Release

Propose tagging v0.1.0a6 after merge (setuptools-scm; move the CHANGELOG [Unreleased] entry under v0.1.0a6); execution-specs then bumps ethereum-hive>=0.1.0a5,<1.0.0 → >=0.1.0a6,<1.0.0.