fix(analyzer): tolerate malformed qpc_freq / truncated rows instead of crashing

[mledour]

Finding #8 (code review): analyzer crashes where the C++ merge degrades

analyze.py parsed the qpc_freq header and every data row with bare int(). So a non-integer # qpc_freq= value (1e7, 10 MHz) or a half-flushed / short data row (a crash mid-write) raised an uncaught ValueError/IndexError and aborted the whole analyzer — whereas the in-DLL C++ merge tolerates both (StrictStoll falls back to the 10 MHz default; ReadFrameCsv/ReadGpuCsv skip a row whose four fields don't all extract). The two documented-equivalent paths thus diverged on corrupt input: one crashes, the other degrades.

Fix — mirror the C++ behavior

• qpc_freq: fall back to DEFAULT_QPC_FREQ (10 MHz) on a parse failure (matches StrictStoll).
• data rows: skip via a _parse_int4 helper that returns None on ValueError/IndexError, with a stderr count of how many were dropped (matches the C++ row-skip).

The loud header-mismatch guard (wrong column header → SystemExit) is kept — that catches "you passed the merged CSV by mistake", a different, deliberate case.

Verification (ran locally)

Two new tests — a malformed qpc_freq header and a truncated 3-field data row — now both produce output (rc 0) instead of a traceback, with the valid frames still merged:

$ python3 -m pytest scripts/test_analyze.py -q
19 passed

CI also runs this suite. No lines >99 cols.

🤖 Generated with Claude Code