nx_galx_repair.nx

nx_galx_repair.nx -- recover the index entries the builder silently drops. FOUND BY MEASURING THE WHOLE POPULATION, NOT A SAMPLE. nx_galx_audit emitted two counters for the same set -- `header_entries 235,555` and `occupied 233,312` -- and the 2,243 gap reconciled EXACTLY with the TSV rows whose char-69 is not a tab. ★★★★★★ TWO COUNTERS OF THE SAME POPULATION ARE A FREE AUDIT; ONE COUNTER IS A CLAIM. The builder counts every newline into n_entries but only PLACES rows it can frame, so 2,243 images were counted, dropped, and never servable -- with no error, and a header that said they were there. The classes, measured (they sum to the 2,243, no residual): 1,171 no cid at all — 622 empty lines + 549 lines starting with a tab (path, no cid) 619 cid field 138 chars = TWO 69-char cids run together 327/106/10 cid field 74-76 chars 5 207 chars = THREE cids run together 138 = 2x69 and 207 = 3x69 is the signature of a writer that never terminated its line: the row was flushed as a bare `<cid>` with no tab, no path and no newline, and the next row landed on it. ⇒ THE PATH FOR THOSE CIDS IS NOT IN THIS FILE AT ALL. It cannot be parsed back out; it has to be recovered from somewhere else. Two somewhere-elses exist: B) an orphaned cid whose blob is still on disk at the conventional <blobdir><cid>.png C) an orphaned PATH whose cid can be RECOMPUTED from the image's own GENREC factors ★ A ROW THAT CANNOT BE PARSED IS NOT NECESSARILY DATA THAT IS LOST — ASK WHAT ELSE KNOWS IT. Usage: nx_galx_repair <tsv> <idx> <blob> <blobdir> [--apply] Dry run by default. The insert is idempotent, so --apply is safe to repeat. license_tier: ORIGINAL