refactor: dedupe stats reduction (#15) + name the data-loss guard (#14)

[mledour]

Findings #15 + #14 (code review): cleanup

#15 — ComputeStats mean/min/max triplet ×3 (+ analyze.py mirror)

ComputeStats repeated the "Kahan-Neumaier mean + min_element + max_element" triplet three times (ms / pct / gpu), and analyze.py mirrored it. Extracted one Summarize() (C++) / _summarize() (Python) so the reduction has a single source — a change to the summation strategy can no longer drift between the three stat groups or across the two languages (the exact "fixed in one branch, missed in the others" risk the review called out).

Pure refactor — the merged-CSV header is byte-identical. The 19-test analyzer suite (which checks the header exactly) still passes:

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

#14 — name the data-loss invariant

The "did this session record any frames?" check was an inline g_frameCounter == 0 with a long comment. Extracted SessionRecordedFrames() as the single definition, used by MergeIntoOutput's zero-frame guard, with the rationale consolidated there (FrameCsvSink's lazy-open documented as the per-side counterpart).

Scope note: #3 already removed the worst part of #14 — the removeStale call sites — by switching the merge to write-temp-then-atomic-rename. So what remained was this naming/comment consolidation. The deeper "right altitude" fix (#13-adjacent) — per-session output filenames so a no-op session physically can't overwrite a prior one — stays deliberately deferred (keeps one file per PID; noted in commit b643964).

3 files, +53/−35; no lines over the limit.

🤖 Generated with Claude Code