Under start.bat the server's stdout/stderr are pipes into PowerShell's
Tee-Object, so Python encodes with the ANSI code page instead of the
wide console API. cp1251 (Russian Windows) cannot encode the banner's
box-drawing rule (U+2500), and that print killed the server with
UnicodeEncodeError before it could listen.
- _guard_console_streams() (the first statement of main()) sets
errors="replace" on both streams: an unencodable character prints as
'?' and can never raise.
- _console_rule() picks the widest banner rule its stream can encode -
U+2500 where the code page has it (utf-8, cp437), ASCII dashes
otherwise (cp1251, ascii).
- The seven text-layer trace strings are ASCII now: the APDU arrows
(U+2192) and the INIT/CAP-COMPAT em dashes (U+2014 - cp437 cannot
encode them either), so the traces stay readable on any code page.
The raw os.write paths (_log_stdout, the APDU tracer) bypass the text
layer, cannot crash, and were left as-is.
Tests: tests/test_console_encoding.py - the strict-cp1251 control, the
rule's fallback/pretty cases, the guarded banner+traces, a stream
without reconfigure, the main() source pin, and an AST pin that no
print/stdout/stderr literal in the package is non-ASCII.
tests/test_startup_cardless.py's CARDLESS_LINE follows the new ASCII
line.
Verified with PYTHONUNBUFFERED=1 PYTHONIOENCODING=cp1251: the pre-fix
code died at the banner, the fixed code reaches the listening banner
and serves a command (CMD: -> OK, no encode errors).
930 frontend + 952 Python; sw cache simple-v453.
__main__ built a plain HTTPServer (single-threaded): during a RAM install
POST every other request waited in the accept queue, including the 2 s
/api/status poll that feeds the operation modal - its bar and step
headline never moved and the whole result appeared only at the end (live
2026-09-29).
_build_http_server() now returns a ThreadingHTTPServer (daemon threads).
Card access is unchanged: it is serialized by _CARD_LOCK, not by the
server - every POST is blanket-wrapped in do_POST, every /api/ GET except
/api/status and /api/test/status is wrapped in do_GET, and the background
threads (STATUS poll, _bip_data_available, test runner) take the lock;
the audit found no card-access holes. do_GET also lets non-/api (static
PWA) paths bypass the lock, so page assets stay answerable during long
operations.
Regression test tests/test_http_concurrency.py: a fast request must
complete while a slow one is in flight (fails on HTTPServer with exactly
the reported symptom: 1.01 s vs 0.20 s); test_startup_cardless patches
the new factory.
sw cache simple-v312; tests: 519 Python / 674 Node.
- __main__ catches NoCardError from the startup init (fast and stock paths,
and the stock fallback) and reports it with one line - 'INIT: no card in
the reader - server ready; insert a card or press Equip' - without a
traceback and without the stock-pysim fallback. Any other init failure
still falls back with its traceback.
- The cardless PysimApp construction no longer prints pySim's 'Waiting for
card...' or pySim-shell's 'pySim-shell not equipped!': a new _LineFilter
(server.py) drops whole matching lines while proxying encoding/isatty/
fileno (the cmd2 Rich console probes them on its stdout).
- Removed the now-duplicate 'INIT: card not initialized' tail line and the
cat_cla assignment when no card was detected (scc exists but is cardless).
- Tests: tests/test_startup_cardless.py (line filter semantics incl. partial
lines, one-line cardless report, no fallback call, real failures keep the
fallback + traceback, muted PysimApp construction). Live-verified against
an empty SCR 3310: two init lines, no traceback, API serves 2.2.17.
- Docs: READMEs quick start, help EN+RU (Card reader), AGENTS (equip flow).
SW cache otaman-v184.
catarrh