fix(input): discard unrecognised CSI sequences instead of typing them

[basil-k-aji-dev]

Fixes #5

TerminalEventParser recognised only a small set of CSI sequences. SGR colour sequences — \x1b[31m, \x1b[0m — matched none of them, and failed two different ways depending on timing.

Typed at the remote host. After ESCAPE_DELAY the ESC fallback did del self.buffer[0], removing only the ESC byte and emitting an ESCAPE key. The remaining [31m was then parsed as ordinary characters, so a single colour sequence produced five keystrokes:

ESCAPE, '[', '3', '1', 'm'

Ended the session. Before the delay elapsed the loop breaks and left the sequence buffered, stalling every byte behind it. The read loop in session/direct.py only calls flush() when select() times out, so during a paste feed() runs back to back and the residue is never cleared. Once it crossed the 8192 guard:

raise ValueError("terminal input sequence exceeds 8192 bytes")

That propagates out of _input_worker, which forwards it via self._controls.put(exc), and the main loop re-raises it with raise control. About 8 KB of coloured text is enough — git diff output, coloured logs, anything copied from another terminal.

Change

CSI_ANY_RE matches a complete CSI sequence (ESC [, parameter bytes 0x30–0x3f, intermediate bytes 0x20–0x2f, final byte 0x40–0x7e). It is consulted only after every specific handler has declined, so a sequence this parser does not act on is discarded rather than buffered or retyped.

It sits after the existing incomplete-sequence check, so partial sequences still buffer and split reads keep working — feed(b"\x1b[12;") then feed(b"40R") still yields one cursor report.

Tests

Four added to tests/test_input.py. The first three fail on main:

FAILED test_unrecognised_csi_is_discarded_not_typed
FAILED test_coloured_text_yields_only_its_visible_characters
FAILED test_bulk_coloured_input_does_not_exhaust_the_buffer

with the first showing exactly the spurious keystrokes:

- [KeyEvent(action=2, modifiers=0, key_code=2, unicode=0),
-  KeyEvent(action=2, modifiers=0, key_code=0, unicode=91),    # '['
-  KeyEvent(action=2, modifiers=0, key_code=0, unicode=51),    # '3'
-  KeyEvent(action=2, modifiers=0, key_code=0, unicode=49),    # '1'
-  KeyEvent(action=2, modifiers=0, key_code=0, unicode=109)]   # 'm'

The fourth guards against regressing arrow keys, cursor reports and partial-sequence buffering.

pytest -q tests/ — 92 passed (88 before). ruff check clean on both changed files. Python 3.13.3, Linux aarch64.

Note

The 8192 guard is left as it is. It still measures the whole buffer before parsing rather than the residue after it, so it can in principle fire on input that is merely unconsumed rather than on one oversized sequence — but with unknown CSI sequences no longer accumulating, I could not reach it with realistic input. Tightening it to check after the parse loop would match the wording of its message; happy to include that here if you would prefer.