Normalize invalid digest inputs

[biru-codeastromer]

Summary

• validate runtime types at the exported sequence and dataset digest boundaries
• reject mixed mapping keys, malformed pairs, boolean and oversized counts, and non-string digests as SS012
• preserve valid digest ordering and byte framing

Validation

• make release-gate
• 107 passed, 1 optional skip
• 92.07% coverage
• clean wheel and source installation
• dependency, text, repository, example, and full-history Gitleaks audits passed

This PR is intentionally left unmerged pending maintainer review.

[biru-codeastromer]

The whitespace-encoding blocker is fixed at the new head. Sequence and dataset digests now require exactly 64 ASCII hexadecimal characters before decoding; space, newline, and tab spellings all return SS012. Exact local gate: 110 passed, 1 optional skip, 92.05% coverage, clean installs, dependency/repository/text audits, example, and full-history Gitleaks all green.

[biru-codeastromer]

The streaming-stack interaction is fixed at the new head. sequence_digest now accepts one-pass iterables, explicitly rejects text and byte containers, and applies the exact 64-character grammar to every yielded value. The generator regression passes. Fresh local gate: 111 passed, 1 optional skip, 92.05% coverage, with clean installs and all audits green. Combined #23 byte-equivalence and bounded-memory validation follows.

[biru-codeastromer]

Combined validation of #22, #23, and this exact head passed: 149 tests, 1 optional skip, 92.53% coverage, including streaming/in-memory byte equivalence, bounded-peak coverage, Parquet, signatures, clean 0.4.0 installation, all audits, example, and Gitleaks. Synthetic 10,000 and 50,000 record runs used 710,840 and 712,942 peak Python bytes respectively, confirming that the generator path remains bounded.

[biru-codeastromer]

Superseded by merged #41, which preserves this change in the protected v0.4.0 release stack.