fix(mcp): derive streamable HTTP retry backoff from backoffs taken

[rxits]

Summary

MCPServerStreamableHttp.call_tool replaces the shared retry loop with its own, and takes the backoff exponent from retries_used — a retry budget counter rather than a count of backoffs already taken. The two disagree in both directions:

• retries_used is only advanced at the top of the loop when max_retry_attempts != -1. So with max_retry_attempts = -1 ("retry indefinitely") the generic except Exception path evaluates retry_backoff_seconds_base * 2**0 on every pass and the delay never grows — an unreachable server is retried at a flat interval forever.
• One isolated-session retry charges retries_used twice, once at the top of the loop and once in the _IsolatedSessionRetryFailed handler. With a finite budget the delay then quadruples per sleep: 1s, 4s, 16s instead of 1s, 2s, 4s.

.agents/references/local-mcp-server-lifecycle.md states that generic list_tools() / call_tool() retries "use the configured attempt count and backoff", and _run_with_retries — the schedule every other retry path in this file follows — is base * 2 ** (attempts - 1) over actual attempts. This change makes the streamable HTTP override agree with it by counting backoffs in a dedicated local and using that as the exponent. Budget accounting, attempt counts, and isolated-session retry eligibility are unchanged.

Test plan

Two tests added to tests/mcp/test_client_session_retries.py, both patching asyncio.sleep to record the delay sequence:

• test_streamable_http_backoff_grows_with_unlimited_retries — max_retry_attempts = -1, expects [1.0, 2.0, 4.0]. On unmodified main it records [1.0, 1.0, 1.0].
• test_streamable_http_backoff_matches_generic_schedule_on_isolated_retry — finite budget with an isolated-session retry, expects [1.0, 2.0, 4.0]. On unmodified main it records [1.0, 4.0, 16.0].

Both were confirmed failing against unmodified main before the fix and passing after it.

.agents/skills/code-change-verification/scripts/run.sh: make format, make lint and make typecheck pass. make tests reports a single failure, tests/test_run_step_execution.py::test_multiple_tool_calls_raise_late_fatal_sibling_exception_after_cancellation[20], which is pre-existing and unrelated to this change: it reproduces identically on unmodified main at 19e364c under the same pytest -n auto --dist worksteal -m "not serial" invocation, and passes when that test is run on its own. I have left the "all verification steps pass" box unchecked for that reason.

Issue number

N/A

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR