[iliaal]

Summary

• Remove the unused dictionary-credential flag from the private token classifier, along with its tuple wrapper and per-token bookkeeping.
• Keep the existing credential-anchor decision after Roman-initial filtering as the single source of truth. No public API or parsing behavior changes.
• Add five regression cases covering Roman initials before/after credentials in the same or separate comma segments, and an initial that must not anchor an unknown candidate.

Validation

• PHP 8.4 locally: 1,197 tests / 3,103 assertions passed (PHPUnit 12.5.23).
• All five exact-head CI checks passed: PHP 8.3 / 8.4 / 8.5 tests (including Composer audit), static analysis (including the PHP 8.5 compatibility gate), and code style.
• A 4,096-case differential comparison against master produced identical serialized outputs.
• PHP syntax checks, empty PHPStan baseline guard, PHP 8.5 banned-call guard, README guard, and git diff --check passed.
• Verified all 84 remote file paths, modes, and blob hashes against the tested local tree before opening this PR; recursive tree was not truncated.

This is a dead-code/contract simplification; no throughput improvement is claimed.