GeezerVison

Follow-up to #79.

Problem

Server._query_output() decodes each service response on its own with read_sized_string(). The service fills text output responses by size, not by character. With a multi-byte encoding, a response can therefore end in the middle of a character. Neither half then decodes on its own:

• encoding_errors='replace': each half becomes U+FFFD, so a 2-byte é comes back as two replacement characters. The text is corrupted without any error.
• encoding_errors='strict' (the default): UnicodeDecodeError is raised part-way through the stream, and output reading stops.

Trace output is where this shows up, because it carries the SQL text of statements.

Fix

Responses are decoded through codecs.getincrementaldecoder(self.encoding)(self.encoding_errors). The decoder keeps the bytes of an unfinished character until the next response completes it.

• It is created on the first response and reset in _reset_output() together with the rest of the output state. Each service call starts clean and picks up the current encoding.
• It is flushed (final=True) at the end of output and at the end of each LINE-mode line. Bytes of a character that never completes still go to the error handler instead of disappearing, so genuinely invalid input behaves as it does today.

read_bytes() is read_sized_string() without the .decode(), so nothing else changes.

Evidence (live, Firebird 5.0.4 on localhost, Windows)

The probe uses the public API only:

• A trace session on employee with log_statement_finish = true.
• Four threads run select 'ação çé ÁÉÍÓÚ ãõ' as t from rdb$database /* same text x20 */ over UTF8 attachments for 3 s.
• The output is read with mode = TO_EOF, query_timeout = 1, readline() and encoding = 'utf-8'.

The replacement characters always come in pairs. Each pair is one 2-byte character cut at a response boundary, for example ação ç followed by two U+FFFD where é was.

Tests

• New test_text_output_multibyte_character_split_across_responses. It splits ç across two responses and is parametrized over LINE/TO_EOF, with and without a timeout between the halves, and over strict/replace.
  • On master, all 8 cases fail: UnicodeDecodeError, or two U+FFFD in place of ç.
  • With the fix, all 8 pass.
• New test_text_output_incomplete_character_at_end_uses_error_handler. A character that never completes still yields U+FFFD instead of being dropped. It passes on master too; it guards the flush.
• The output_response() test helper now also accepts bytes.
• tests/test_server.py against a live server: 56 passed, 1 skipped. Master gives 46 passed, 1 skipped.
• Full suite: 173 passed, 16 failed, 2 skipped. Master gives 163 passed with the same 16 failures (identical test IDs). On my machine those are all environmental:
  • 8 in test_fb4.py: the server raises an integer-overflow DatabaseError.
  • 8 in test_provider.py: OSError: [WinError 193].
• ruff check gives identical findings on both changed files before and after.