client: drop SelectiveHTTP2Transport, use HTTP/1.1 only

[ggprior]

Summary

• Drop SelectiveHTTP2Transport and the HTTPTransport import — ServiceClient.httpx_client now uses HTTP/1.1 everywhere.
• Drop the [http2] extra from the httpx dependency so h2 is no longer pulled in transitively.
• Inline comment explains why in case anyone is tempted to reintroduce HTTP/2 in the future.

Why

The transport opted /tabpfn/fit and /tabpfn/predict into HTTP/2 while leaving everything else on HTTP/1.1. Both are long-running unary POSTs (thinking-mode fits keep the stream open for 5–15 min) and gain nothing measurable from HTTP/2's multiplexing or HPACK — one request per fit, multipart body dominated by raw bytes.

What HTTP/2 did buy us: a state-machine race.

The h2 library treats a PING frame received while ConnectionState == CLOSED as a protocol violation (RFC 7540 §6.7 — PINGs only make sense while the connection is open). During long thinking fits, intermediate keepalive PINGs (Cloud Run's LB sends them on long-lived streams) regularly land on the client after the connection has been reaped by some hop, and httpx surfaces this as:

httpx.LocalProtocolError: Invalid input ConnectionInputs.RECV_PING in state ConnectionState.CLOSED

That exception class is not in either _fit's or _predict's @backoff.on_exception retry tuple — the tuple covers RemoteProtocolError (peer sent garbage) but not LocalProtocolError (our state machine got wrong-footed by a benign-but-out-of-order frame). So instead of being retried silently, the fit fails hard with a confusing error.

Reproduction / evidence

A 135-fit TabArena-Medium sweep against api.priorlabs.ai (15 datasets × 3 folds × 3 modes: plain / thinking_medium / thinking_high) ate 6 LocalProtocolErrors across thinking_medium + thinking_high:

{"error": "Invalid input ConnectionInputs.RECV_PING in state ConnectionState.CLOSED",
 "traceback": "...httpcore/_sync/http2.py:185...raise LocalProtocolError(exc)..."}

Re-running the same matrix locally with httpx_client monkey-patched to drop the selective HTTP/2 transport cleared all 5 missing thinking_high cells with zero protocol errors — confirming the race disappears.

Alternative considered

Adding httpx.LocalProtocolError to the retry tuple would also mask the issue, but it's strictly the worse fix:

• still leaves HTTP/2 enabled with no upside
• one retry per fit, ~5–15 min wasted on the doomed attempt
• the h2 state machine can fail in other ways too (less common, but still in scope)

Dropping HTTP/2 entirely eliminates the class of bugs at the cost of zero observable functionality.

Test plan

• pytest tests/unit/test_client.py — 22/22 pass (unchanged from main).
• Local rerun of 5 previously-failing thinking_high TabArena-Medium tasks under the patched client — 0 protocol errors, all 5 complete cleanly.

[gemini-code-assist]

Code Review

This pull request disables HTTP/2 support by removing the SelectiveHTTP2Transport class and reverting to HTTP/1.1 for all requests. This change addresses a protocol violation issue where long-running requests would fail due to keepalive PINGs from the load balancer being received after a connection was closed. The httpx[http2] dependency has also been simplified to httpx. I have no feedback to provide.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[ggprior]

Post-patch validation:

Re-ran the failed-row subset of the TabArena-Medium sweep (thinking_high, 45 fits over 5h 56m on api.priorlabs.ai) against this branch's patched SDK:

	HTTP/2 (main)	HTTP/1.1 (this PR)
fits attempted	45	45
fits completed	40/45	45/45
LocalProtocolError (RECV_PING in CLOSED)	5	0
backoff retries needed	0 (LocalProtocolError bypasses the retry tuple)	0 (no transient issues at all)
total wallclock	6h 31m	5h 56m

Zero protocol errors across 5h 56m of long-running thinking fits — the 4.4% intermittent failure rate from the HTTP/2 PING race is gone.