fix(#1657): resolve three independent Windows papercuts

[raman118]

1. The Accented Path Mojibake Massacre

• The Annoyance: If you had files named with beautiful accents (like Confirmé.md or Expérience.py ), Windows would choke on them. They’d get written
into manifest.json and GRAPH_REPORT.md looking like total garbage ( Confirm or Expérience ).
• The Root Cause: Standard Python Path.read_text() and .write_text() calls without explicit encoding fall back to the system locale default. On
Windows, that means cp1252 (Windows-1252). When UTF-8 characters were passed through, Windows decoded them using cp1252 , causing a classic encoding
leak.
• The Fix: We went into load_manifest and save_manifest inside graphify/detect.py and wrapped the read/write boundaries in clean open(...,
encoding="utf-8") blocks. Crucially, we also added ensure_ascii=False to json.dump() , forcing the JSON writer to save raw UTF-8 characters to the
file instead of converting them into escaped ASCII sequences ( \u00e9 ).
• The Safety Net: We added test_manifest_and_report_preserve_accented_paths to check that accented names round-trip perfectly byte-for-byte in both
the manifest and the generated markdown report.

2. The Windows Console Encoding Meltdown

• The Annoyance: Running graphify query on Windows without manually setting PYTHONIOENCODING=utf-8 would lead to ugly console crashes (
UnicodeEncodeError / UnicodeDecodeError ) whenever non-ASCII characters were returned in the output.
• The Root Cause: The default Windows command console doesn’t use UTF-8 by default. When Python tried to push non-ASCII characters to standard output,
it threw a fit.
• The Fix: At the absolute top of the CLI entry point ( graphify/main.py ), before any output gets printed, we added a reconfiguration guard:
if hasattr(sys.stdout, "reconfigure"):
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
This forces both stdout and stderr to speak UTF-8, using replacement characters if something is truly mangled, without breaking piped/redirected
output.
• The Safety Net: We added test_query_cli_non_ascii_no_encoding_error in test_query_cli.py . It spawns a subprocess running the query CLI without
PYTHONIOENCODING to simulate the raw Windows environment, proving it no longer crashes.

3. The Import Cycles Noise Pollution

• The Annoyance: When running graphify on a corpus composed mostly of documents (like a folder full of markdown papers), the report would print a
section called ## Import Cycles . Because import cycles are a code-only concept, this was pure noise, usually returning "None detected" or parsing
markdown paths incorrectly.
• The Root Cause: The report generator was blindly appending the Import Cycles section without checking whether the corpus actually contained code.
• The Fix: In graphify/report.py , we added a quick scan to count the number of nodes in the graph that represent code. If has_code_nodes is false,
we skip building the ## Import Cycles section entirely.
• The Safety Net: We wrote two new tests: test_report_omits_import_cycles_on_document_only_corpus and
test_report_includes_import_cycles_on_code_heavy_corpus . Now we explicitly assert that the cycles section is omitted on markdown/document-only trees,
while remaining perfectly intact on code-heavy repos.