numeric_results

Numeric parsing results

Measured evidence

• The outlined exponent fallback reduced the exponent microbenchmark from the original fallback cost (about 6.5x the plain parser) to 30.1--30.7 ms; the prior exponent path measured about 53.2 ms.
• The integrated outlined helper kept plain CSV parsing near 22 ms, matching the baseline in the paired measurement.
• Generic integer parsing retained a bounded packed path only where its width/length gate proves the magnitude is in range. The measured 20M-field results are recorded in integer_swar_results.md.
• The bounded direct Float32 subset improved its matching short input by 2--7%, but its intentional fallback was about 35% slower. It is rejected as the default path; see float32_direct_audit.md.

Source inspection

The strict long-decimal route remains safe by construction: on the twentieth digit, the scanner calls the exponent-specialized fallback, which in turn uses the checked borrowed _atof conversion. It does not retain a partially accumulated mantissa. This covers ambiguous and more-than-19-digit decimals with the standard conversion's existing behavior.

The packed integer validator operates on raw UInt8 words and rejects a byte outside ASCII 0x30..0x39 through its high-bit range test before digit reduction. Non-ASCII UTF-8 bytes therefore reject as non-decimal rather than being interpreted as digits. The implementation deliberately has a little-endian compile-time requirement.

No additional strict-grammar, rounding, or overflow defect was identified by source inspection. This is inspection evidence only: exact grammar and rounding confidence comes from the existing float reference tests (including the exponent corpus) and exhaustive packed-integer byte tests, rather than from a new benchmark or test run in this task.