a press measured, under a terminal, against the budget

[tamnd]

The shell has stopped a running statement on Ctrl-C since the interrupt landed, and the C ABI has answered zu_conn_interrupt for as long, but neither said how long it took. Under a pseudoterminal the shell took forty milliseconds against a budget of fifty, which is inside it and too close to it to be a number anybody would want to defend on a busy machine.

The reason is the watcher's sleep, and a press pays it twice: once for the wake that reads the flag the handler set, and once for the wake that finds the statement over, because the loop only looks between two sleeps. Twenty milliseconds each makes forty. Five makes ten, and costs a compare and a sleep two hundred times a second while a statement is running, which is nothing beside the statement it is waiting for. The progress line does not change either way, since it is written only when the tenth of a second or the row count it prints has changed.

crates/zu-cli/tests/press.rs is the measurement. The shell only watches for a press where standard input is a terminal, so the test gives it one: a pseudoterminal from posix_openpt, the shell on the far side of it with the editor and the progress line it would have under a person. It types the pairwise statement over twelve thousand people, waits for the progress line so that the press lands on a statement inside the executor rather than one still being parsed, sends the signal a terminal driver sends on Ctrl-C, and reads the interrupted line back at eleven milliseconds. Then it asks the same session another statement and gets the count, because a stop that took the session with it would be a different bug wearing this one's answer.

The press is a SIGINT to the child rather than a ^C byte into the terminal, because that byte only becomes a signal once the driver's signal characters are on and the child is the foreground process group of a controlling terminal, which is session bookkeeping the test would do to arrive at the same signal. Unix only, since Windows has neither this pseudoterminal nor this signal, and the console equivalent is a mechanism the shell does not use yet.

The C ABI's stop test grew the same assertion around the call it already made, timing the ask against the return of zu_query, since there the ask is a call and not a signal.

docs/10-api-and-tooling.md says the poll rate it now has and where the three gates live.

Local gates: cargo fmt --all --check, cargo clippy --workspace --all-targets --all-features -- -D warnings, and cargo test --workspace --all-features, green.

Part of DX2, #168: Ctrl-C interrupts a 10-second query within 50 ms in the shell, in Python, and through the ABI. The Python half is tamnd/zu-python#8, merged.