GeezerVison

Thrift ExecuteStatement is excluded from retryableThriftMethods, so any transport failure fails the query on the first attempt. Retrying it unconditionally (as the conformance spec and the reference ADBC PR do) can run an INSERT/MERGE/DDL twice. I reproduced it locally on Node 24 with the driver's own node-fetch and keep-alive agent. A connection that is closed after the peer read the full request (what the conformance proxy's close_connection does) and a stale keep-alive socket whose request was never read look exactly the same to the client: ECONNRESET "socket hang up", on a connected, reused socket, with the request fully written. So Node can't prove those requests were never delivered.

This PR retries ExecuteStatement only when the connection fails before it is established, which is the case where we can prove nothing was sent: ECONNREFUSED, ENOTFOUND, EAI_AGAIN. It uses the existing retryMaxAttempts / retriesTimeout / backoff settings. HTTP status codes and failures on a connected socket are still not retried. When retries run out, the original connection error is thrown instead of a RetryError, so callers keep the error they get today. Other methods are unchanged.

Verification:

• New unit tests use a local server and an agent that refuses connects on demand. A refused connect is retried and succeeds. A close after the request was sent, and a 503, both surface after exactly one attempt. OpenSession is still not retried. Idempotent methods still retry. Full unit suite: 1548 passing. Lint, prettier, type-check and build are clean.
• Live, via a local CONNECT proxy in front of the e2e warehouse:
  • Proxy down, so the connect is refused: origin/main fails in 1 ms with ECONNREFUSED. This branch retries after the 2 s backoff and returns the row.
  • Proxy forwards the ExecuteStatement bytes, then drops the tunnel before the response: both builds surface the same ECONNRESET after one attempt.
  • Thrift e2e specs (operation_error_codes, timeouts, query_parameters, proxy, iterators): 17 passing. useKernel: true SELECT 1 works.
• Perf: the success path now builds the new policy instead of a NullRetryPolicy, about +0.2 µs per ExecuteStatement (median 88 ns vs 292 ns, 10 runs each). Loopback ExecuteStatement round-trip median: 190.5 µs (main) vs 197.7 µs (branch), within that benchmark's noise. Live ExecuteStatement median: 106.5 ms (main, 5 runs) vs 107.7 ms (branch, 6 runs).

Not addressed here:

• The conformance scenario itself (connection closed after the request was sent) is intentionally still not retried, so the ERRORRECOV-013 Thrift xfail stays red. Retrying it needs server-side idempotency or a product decision to accept duplicate execution.
• [sea]: the kernel not retrying a connection close on ExecuteStatement is in databricks/databricks-sql-kernel (databricks/databricks-sql-kernel#185), not in this repo.

Addresses #458

This pull request and its description were written by Isaac.