test: eliminate PTY echo race in Python interactive-exec E2E

[matthewgrossman]

User Story

Matt (MG) needs reliable OpenShell branch E2E checks while qualifying PR #4066. The Docker Python lane failed its interactive-exec TTY assertion at an unchanged SHA after an earlier successful run, and Matt requested an investigation and a focused draft fix.

Problem Statement

e2e/python/test_sandbox_api.py::test_sandbox_interactive_exec_honors_tty can fail its contiguous TTT assertion even when all three descriptors are terminals, stdin is consumed, both output sentinels arrive, and the command exits successfully. The shell writes each TTY flag separately while the request generator immediately sends stdin; terminal echo can arrive between flag writes.

Attempt 2 failed on October 1, 2026, with 1 failed, 59 passed, 81 skipped. Attempt 1 passed at the same SHA (60 passed, 81 skipped). This was a PR-branch failure. Main at 348a1fc62566892ba5530ac9f35acc983e4d9324 retains the same relevant test/runtime source.

Impact / Why This Matters

The failure prevents repeatable qualification of an unrelated Nix download-recovery change. Rerunning CI consumes time and can produce another false failure without improving coverage.

Acceptance Criteria

• PTY input echo cannot split the TTY readiness assertion.
• The test still rejects lost TTY support, lost consumed stdin, and missing output.
• Non-TTY stdout/stderr separation checks remain intact.
• Readiness waiting is bounded and request iterators unblock on early exit or RPC failure.
• The focused test and Docker Python E2E lane pass against OpenShell.

Reproduction Steps

1. Run OpenShell's Docker Python E2E lane with mise run e2e:python.
2. Repeat the interactive-exec TTY test against the Docker-backed gateway. Its input generator immediately submits streamed-stdin-sentinel\n after the exec start message.
3. When PTY echo is scheduled between the second and third shell flag writes, the test fails despite successful terminal execution. This is intermittent; the two linked CI attempts provide the observed reproduction.

Environment

• OpenShell SHA: 6c4bb7614af2d86783e8a887b8915ef1520e94df (draft PR fix(ci): retry Nix shell and app dependency preparation #4066).
• Deployment: OpenShell Docker-backed gateway and Linux sandbox, Branch E2E Checks.
• Python: 3.14.5; pytest: 9.0.2.

Logs

assert b'TTT' in (b'TTstreamed-stdin-sentinel\r\nT\r\nstdin:streamed-stdin-sentinel\r\nstdout-sentinel\r\nstderr-sentinel\r\n' + b'')

Proposed Fix

Wait for a complete output readiness marker before streaming stdin, preserving the actual TTY and pipe assertions. A controlled real-PTY/Python-gRPC reproduction produced the exact original failure; the proposed handshake passes the same ordering and rejects deliberate regressions. No production SDK/runtime changes are proposed.

Related testing strategy tracker: #3954. This issue covers only the interactive-exec TTY flake and should be closed independently.

[matthewgrossman]

I specifically told chatgpt to make a draft PR and not an issue, smh