fix(copyright): bound and normalize author candidates

[mstykow]

Summary

• Stop author candidates at grammatical lowercase prose and change-action boundaries while preserving conjoined and handle-bearing rosters.
• Normalize quoted names and punctuated obfuscated contacts idempotently across repeated refinement.
• Reject identity templates and subject-role object phrases that are not file authorship.

Issues

• Covers: follow-up findings from False-positives when comparing provenant with scancode-toolkit output #1391

Scope and exclusions

• Included: generic final author refinement and grammar-context cleanup.
• Explicit exclusions: repository-specific values, filename checks, and package-party extraction.

How to verify

• Scan Perl 5.38.4 with provenant scan --copyright --json out.json perl-5.38.4. Relative to fix(copyright): recover bounded POD contributor rosters #1412, seven malformed or false author values disappear and six supported values appear: three clean replacements (Larry Wall, Nicholas Clark, Paul Marquess), two newly recovered Benchmark contributors, and one structured obfuscated-contact author. Quoted duplicates, a command template, and an obituary object phrase are removed. Copyright and holder output is unchanged.
• Scan InfoZIP 3.0 as a control. Its only author delta is two occurrences of Greg Roelofs newt (at) pobox.com normalized to the idempotent Greg Roelofs newt at pobox.com; copyright and holder output is unchanged.
• Run cargo test --lib copyright::detector, cargo test --test copyright_golden --features golden-tests, and cargo test --test post_processing_golden --profile ci-release --features golden-tests.

Intentional differences from Python

• ScanCode's token grammar avoids some of these Provenant overrun shapes but misses both Benchmark contributors, and it reduces the structured obfuscated-contact author to Masaaki Goshima. Provenant retains the supported identity and contact while enforcing explicit grammar boundaries.

Follow-up work

• Created or intentionally deferred: source-header credit recovery remains in the next stacked branch.

Expected-output fixture changes

• Files changed: none.
• Why the new expected output is correct: focused unit and end-to-end tests cover prose/action boundaries, list preservation, quote normalization, templates, subject-role constructions, and repeated obfuscated-contact refinement; both broad golden suites pass unchanged.

[mstykow]

@greptileai review