[TMHSDigital]

Closes #402. Refs #403: this fixes the encoding crash and the stale expect_file. Status-file truncation is deliberately not done, see below.

Problem

• About 284 sequential Blender runs shared one 75-minute job timeout, with no per-run limit, so one hang ate the whole job.
• The status record was only written after wait(), so the summary never named the example that hung.
• On Windows, Blender's UTF-8 output was decoded with cp1252. A byte such as 0x9D raised UnicodeDecodeError in the read loop, which crashed the runner and left Blender running.

Change

• run_example.py:
  • new stream(): a reader thread plus proc.wait(timeout=…);
  • --timeout (default $BDT_SMOKE_TIMEOUT or 900 s);
  • on expiry, kills the whole process group (start_new_session + killpg; taskkill /T on Windows) and records FAIL "timeout after Ns (process group killed)";
  • output decoded as UTF-8 with replacement.
• run_catalog.py:
  • forwards a row's optional timeout;
  • deletes a row's expect_file before running it, so a render left by an earlier local run can't satisfy the check.
• check_smoke_catalog.py: accepts timeout and requires it to be a positive number.
• check_falsifier_targets.py --run (the weekly sweep): a 900 s bound per run; a timeout counts as a mismatch and the sweep continues.
• Not done: truncating --status in run_catalog. CI's canary and template steps append to the same JSONL before the catalog runs, so truncating it would erase their records. CI is unaffected by stale records because RUNNER_TEMP is fresh.

Evidence

Unit tests (tests/smoke/test_harness.py, Python 3.13, Windows): 40 tests OK, 1 skipped (the POSIX grandchild-kill test, which runs on Linux CI). New tests:

• exit code returned;
• a hang is killed and returns None in under 30 s, with output captured;
• grandchild killed (POSIX);
• b'\xe2\x80\x9d' decodes to U+201D;
• a row's timeout is forwarded;
• a stale expect_file no longer satisfies a row.

Falsified:

• With encoding="cp1252", errors="strict" restored, the UTF-8 test fails with UnicodeDecodeError: 'charmap' codec can't decode byte 0x9d.
• With the old untimed proc.wait() restored, the hang test hangs until killed externally, which is the bug.

Live run with real Blender (.scratch binary reporting 5.2.1 LTS):

Run	Result
bmesh-gear via run_example.py	[PASS], exit 0
a script that prints then time.sleep(120), --timeout 5	[FAIL] hang-probe: timeout after 5s (process group killed) after 5 s. Get-Process blender count afterwards: 0. JSONL record proc_exit: null. summarize.py tables it as FAIL with the reason

check_smoke_catalog.py and check_falsifier_targets.py pass.

🤖 Generated with Claude Code

[TMHSDigital]

CI evidence before merge: all checks pass ( 13 pass ). Blender Smoke on PR head: Blender 5.2.2 LTS / Blender 4.5.14 LTS. Summaries (happy path; falsifiers) per leg: 143 passed, 1 skipped, 0 failed;142 passed, 0 skipped, 0 failed;141 passed, 3 skipped, 0 failed;136 passed, 6 skipped, 0 failed;