testing: smoke runner decodes Blender output with the locale codec and reruns reuse stale status/output files #403

Evidence

• Encoding. tests/smoke/run_example.py:~58 uses text=True with no encoding, so on Windows Blender's UTF-8 output is decoded as cp1252.
  • Today's — and → decode without error but come out garbled.
  • A UTF-8 character containing the bytes 0x81, 0x8D, 0x8F, 0x90 or 0x9D (e.g. ”, ←, Í, or a non-ASCII user path) raises UnicodeDecodeError in the read loop. The runner then crashes without a record and leaves Blender running.
• Stale state on local reruns.
  • run_catalog.py appends to the --status JSONL and never truncates it, so old PASS and FAIL records accumulate.
  • It never deletes expect_file (e.g. $OUT/swatch.png) before a run, so a stale render can satisfy the check.
  • CI is unaffected (fresh RUNNER_TEMP); local and Windows contributors are not.

Suggested approach

• Pass encoding="utf-8", errors="replace" to Popen.
• Truncate the status file at catalog start (or namespace it per run).
• Delete each expect_file before its row runs.

Done when

A decode test with b"\xe2\x80\x9d" passes, and the catalog truncates its status file and pre-deletes expected outputs.