[somethingwithproof]

Imports now report failures to their callers and retain source files when ingestion or post-processing fails. The follow-up separates the streaming parser, batch writer, SQL table extractor, page queries and statistics, while keeping the public helper functions available.

• Check import database reads and writes, return nonzero CLI failures, and delete input only after successful inline processing.
• Reset optional parser fields between entries and bound detail batches by row count and serialized size.
• Pass chart and detail filters explicitly to query helpers.
• Share a statistics sample budget across groups and release each scope after writing it. Counts, totals, minima and maxima remain exact; percentile estimates are labeled in the chart view.
• Run the existing integration workflow for stacked PRs and include every extracted helper in the existing coverage check.

Stacked on #60; this PR targets fix/review-hardening so the refactor can be reviewed separately.

Validation: 534 tests pass, including the real database and HTTP suite; CI changed-line coverage is 100% (792/792 lines). Real MariaDB regressions verify retained input and nonzero file-import, post-processing and reprocessing failures. PHP syntax, manifest, translation-template and workflow checks pass. CI passes on PHP 8.2, 8.3 and 8.4.

Five-run isolated PHP benchmarks show statistics peak allocation falling from 30 MiB to 14 MiB, with median time falling from 0.498 s to 0.459 s. Ingestion takes about 9% less time and table extraction is essentially unchanged. Database adapters are stubbed; these are not whole-application latency measurements. Reproduction commands and raw results are in tests/benchmarks/.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
No data about Coverage
0.8% Duplication on New Code

See analysis details on SonarQube Cloud