[khushalsonawat]

Description

test_post_terminal_wire_block_is_bounded_by_the_remaining_flush_window failed once on #273's CI. run_job returned 4, EXIT_TERMINAL_UNDELIVERED, because the terminal's own delivery didn't fit in the test's 0.2 s flush window on a slow runner. The test also asserted the whole run took under 1.5 s, which locally used about 1.3 s of it. Both numbers measured how fast the runner was, not whether the bound holds.

This is a test-only change. The flush-window code in hosted_entrypoint.py is unchanged.

Type of Change

• Tests

Changes

• The window is 2.0 s, not 0.2 s, so the terminal's delivery has room.
• emit_skipped_receipts is replaced by a call that waits on an event that is never set, so it can only end when the window's wait_for cancels it. The stall stands in for the receipt push.
• The test checks the bound the code sets, not elapsed time (62b316b4, after Dileep's review). A spy on wait_for records the timeout passed with the stalled call, and the test asserts 0 < timeout < window. The code passes the remaining window, which is always under the full one, so a late cut and a full-window timeout both fail.
• The run sits under an outer asyncio.wait_for of 30 s, so a broken bound fails with a clear message instead of hanging the suite.
• The 1.5 s wall-clock assertion is gone. So is the check that the second scenario sent no receipt, which can't fail now that the push is never reached. Exit 0, one canceled terminal, the first scenario passed and the pool closed are all still asserted.

The test now always takes about 2 s.

Testing

• tests/harness/test_hosted_entrypoint.py:
  • Python 3.12.14: 139 passed. The changed test takes 2.03 s.
  • Python 3.12.3: 133 passed. The 6 failures are the voice tests, which need livekit; it isn't installed in that venv, and they fail the same way on the base.
• Planted bugs in hosted_entrypoint.py, each run on this test and then restored:

plant	what's broken	result
w1	the wait_for around emit_skipped_receipts removed	fails at the 30 s guard
w2	timeout=None	fails at the 30 s guard
d1	the cut comes 1.8 s late (Dileep's case)	fails at the bound (3.76 against 2.0)
d2	the full window instead of the remaining one	fails at the bound (2.0 against 2.0)

Known limits

• The window still depends on headroom. The window starts when the cancel is seen, before the terminal is sent. If the steps before the terminal take more than 2.0 s, run_job returns 4 again. A planted 2.5 s delay there fails the test; a planted 1.5 s delay passes. Locally those steps take under 0.07 s, so the 2.0 s window is about 30 times what's needed. The source reads time.monotonic() rather than deps.clock, so a test-only change can't control that clock.
• No slow-runner simulation. I didn't run the test under cpulimit or beside a busy loop, to keep the laptop cool. The reviewer measured the same margin directly with the two delay plants above.
• Reviews. One cold review (Lane Q); its Medium and a Low were fixed in 5e35bf80. A fix-only check of 5e35bf80 and 62b316b4, one reviewer: clear, 0 Medium (results in the comments).
• A constant just under the full window would pass the bound check; the code computes the remaining window, and nothing in it would produce such a constant.

Related Issues

Refs: TH-9515

[cdileep23]

Please check the comments.

[khushalsonawat]

Fix-only check of 5e35bf8 (the first review's fix, never checked until now) and 62b316b (Dileep's P2), one reviewer: clear, with 0 High, 0 Medium and 3 Low.

• 5e35bf8: a cut 10 s late fails the old test, but a cut 1.8 s late passes. That is Dileep's P2, and 62b316b replaces the assertion.
• 62b316b: these wrong builds each fail it:
  • a cut 1.8 s late (3.758 against 2.0);
  • the full window instead of the remaining one (2.0 against 2.0);
  • timeout=None, through the 30 s guard.
• The strict < is safe: the timeout came out 42 to 57 ms under the window, far above clock resolution.
• Lows:
  • A constant just under the full window would pass.
  • The comment's wording about None is slightly off.
  • The test could fail on correct code if the whole 2 s window were used before the timeout is computed. About 1.94 s of headroom was measured.
• The test file: 133 passed. The 6 failures are the voice tests that need livekit, which isn't installed on the review machine.

[khushalsonawat]

CI at 62b316b: SDK smoke's clean-install failed one test, and it comes from the base, not this PR.

• This PR's run: test_contract_model.py::test_a_crafted_file_of_block_comments_does_not_hold_up_the_check[a-file-of-openers] took 0.675 s against a 0.5 s bound. Everything else passed: 5957 passed. run
• The base's own run: feat/TH-8804-source-subfolder at 084bca2 fails the same case on its push run. run
• This PR changes only test_hosted_entrypoint.py. The failed job is re-running.