fix(emacs-core): log a routine command-loop signal at debug, not error

[tag-und-nacht]

Fixes #342.

Pressing an arrow key at the end of a line under evil shows "End of line" in the echo area exactly as GNU does, and also logs

ERROR neovm_core::emacs_core::eval: Command loop error: End of line [signal=(end-of-line nil)] condition=end-of-line

on every keystroke, and the same for every C-g. Evil defines end-of-line as an error and pushes it onto debug-ignored-errors (evil-macros.el).

GNU's command loop never logs: cmd_error_internal (src/keyboard.c:1030-1047, emacs-31.0.90) hands the signal to command-error-function and that is the whole handling. The log line is this port's diagnostic, and it ranked every signal the same. GNU ranks signals in one place, the debugger gate skip_debugger (src/eval.c:2146-2180) over debug-ignored-errors (defaults in lisp/bindings.el:1038-1046: beginning-of-line, end-of-line, end-of-buffer, buffer-read-only, user-error, ...), and a quit is routine by construction (signal_quit_p, src/eval.c:2181-2190).

Context::command_error_severity applies that ranking through the existing skip_debugger matcher and the crate's quit predicate, as a closed CommandErrorSeverity { Routine, Diagnostic }. command_loop_2 logs a routine signal at debug and keeps the error level, with condition and payload, for everything else, so a void function in a command still reaches a bug report. Presentation is unchanged: report_command_error still runs first through Lisp's buffer-local command-error-function.

Declared departures from GNU. GNU consults skip_debugger only when the debugger is otherwise wanted (maybe_call_debugger, src/eval.c:2206-2213); the log has no such gate, so the list is evaluated on every uncaught command signal. And the ranking runs inside command_loop_2, GNU's sole recovery owner, where nothing may signal: a string entry in debug-ignored-errors that is not a valid regexp makes the matcher signal invalid-regexp. The classification is therefore infallible and a matcher failure ranks the original error as a diagnostic; a test pins that.

Tests, red first: command_error_severity_follows_gnus_debug_ignored_errors (evil's end-of-line is a diagnostic with an empty list and routine once listed; user-error, a condition inheriting from it, and a regexp entry over the message are routine; quit and minibuffer-quit are routine whatever the list says; void-function stays a diagnostic) and command_error_severity_never_signals_on_a_bad_ignore_list.

Verified on macOS: the two tests, rustfmt, and a GUI run of the release build feeding the real command loop a command that signals end-of-line and one that calls a void function: the first logs at debug only, the second keeps its ERROR line. Not run: Linux and Windows (platform-neutral Rust in neovm-core); clippy on neovm-core's tests stops on a pre-existing never_loop error in the process tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_0117NsCB7AbwgdEqduGF5Kww