Exit quietly when the reader closes the pipe

[ESultanik]

Closes #156

graphtage a.json b.json | head -3 printed roughly 107 lines of chained BrokenPipeError tracebacks on STDERR
once head exited. Piping a large diff into head or a pager is a normal way to read it, so the noise is easy to
hit.

What changed

A single except BrokenPipeError around the diff is not enough, because the failure surfaces at four independent
points:

1. HTMLPrinter.__init__ writes the document header, so --html fails before the diff starts.
2. formatter.print(printer, diff) writes the diff itself.
3. Leaving the with printer: block flushes whatever is buffered.
4. printer.close() in the finally clause flushes again, and HTMLPrinter.close writes </body></html> first.

All four are guarded. The handler inside main() sets a flag instead of returning, because had_edits is bound
only inside the with body and falling through to it after a broken pipe would raise UnboundLocalError.

Catching the exception alone is still not quiet: CPython flushes sys.stdout during interpreter shutdown and
reports the second failure as Exception ignored ... BrokenPipeError. The new silence_broken_pipe() points file
descriptor 1 at os.devnull with os.dup2, which is the remedy the Python documentation recommends for SIGPIPE.
It also lets the cleanup writes that follow succeed instead of raising again.

Exit status

EXIT_BROKEN_PIPE = 141, which is 128 + SIGPIPE, the status a shell reports for a process that a broken pipe
terminated. Python ignores SIGPIPE and raises BrokenPipeError instead, so Graphtage reports the status itself.

EXIT_DIFFERENCES_FOUND (1) would claim the inputs differ, and EXIT_ERROR (2) would claim Graphtage could not
compute the diff. Neither is true when the reader simply stopped listening, and 141 is what a shell with
pipefail already reports for the rest of the pipeline.

graphtage-git-diff

Covered. It writes the repository path to sys.stdout and flushes before delegating to graphtage.__main__.main,
and git runs external diff drivers under a pager, so quitting the pager reaches those writes as well. Its own
writes now return EXIT_BROKEN_PIPE; the delegated diff is covered by the fix in __main__.

Validation

The regression test is test_closing_the_pipe_early_is_quiet in test/test_printer.py. It runs
python -m graphtage on two JSON files whose diff is several times the size of the pipe buffer, reads 64 bytes,
closes the read end, and asserts that STDERR contains neither Traceback nor BrokenPipeError and that the exit
status is 141. It runs for both the plain and the HTML printer.

The test was confirmed to catch the bug: with the guards reverted (keeping only the new constant so the import
resolves), both subtests fail, reproducing exactly the tracebacks from the issue — three chained ones ending at
printer.close() plus the Exception ignored while flushing sys.stdout line. With the guards in place both pass.

Also checked by hand that graphtage c.json d.json | head -3 leaves STDERR free of tracebacks with progress bars
both enabled and disabled, and that set -o pipefail reports 141.

CI run locally on Python 3.14:

• ruff check graphtage test docs bindist — clean
• pytest — 141 passed, 2 subtests passed
• make -C docs html SPHINXOPTS="-W --keep-going" — build succeeded
• uv lock --check — reports the lockfile needs updating, but that is a sandbox artifact (Resolving despite existing lockfile due to addition of global exclude newer); neither pyproject.toml nor uv.lock is touched by
  this branch

🤖 Generated with Claude Code

https://claude.ai/code/session_01GypKU5KdLfs2Cf8kS2TzJa