fix(benchmarks): improve reviewed Rust benchmark snapshots

[mstykow]

Summary

• improve Rust benchmark-review outcomes by fixing scanner behavior around dual-license README handling, weak author noise, bounded extracted-text recovery, and workspace-root Cargo lockfile identity
• keep the compare workflow generic by removing compare-only URL collapsing and avoiding ecosystem-specific package-data equivalence rules in the benchmark compare path
• add reviewed docs/BENCHMARKS.md snapshots for Amanieu/atomic-rs, marshallpierce/rust-base64, and rustcrypto/aeads, then regenerate the benchmark chart and headline stats

Scope and exclusions

• Included:
  • generic scanner fixes for dual-license README notices, weak author noise, binary copyright preservation, extracted-text recovery, and Cargo workspace lockfile identity
  • compare-path cleanup so benchmark review stays generic and literal where it should, instead of hiding URL or package-identity differences
  • reviewed benchmark rows for atomic-rs, rust-base64, and aeads
• Explicit exclusions:
  • no ecosystem-specific compare exception framework was introduced; compare stays generic and the remaining reviewed differences are expressed in benchmark notes instead

Dual-license handling details

• README dual-license notices such as licensed under either of ... at your option and dual-licensed under ... are now treated as true alternative-license signals instead of being flattened into conjunctive output.
• On the detection side, the scanner now combines the substantive README matches with OR semantics when the surrounding notice text clearly describes an alternative-license choice, while still keeping genuinely supplemental material such as warranty-disclaimer fragments separate.
• On the result-shaping side, the branch adds joined-expression selection logic so that when a README produces both singleton fragments like MIT / Apache-2.0 and a covering joined expression like Apache-2.0 OR MIT, Provenant prefers the joined expression instead of surfacing a noisier or accidentally conjunctive end-state.
• The README/license-processing path also stops injecting Apache-specific badge or markdown-hint augmentation into files that already contain one of these explicit dual-license notices, because that hinting was one of the main reasons the original scans drifted toward AND-shaped output.
• Redundant README conjunctive detections are pruned more locally now: the scanner only drops overlapping conjunctive detections that are subsumed by the stronger alternative-license notice, rather than broadly suppressing unrelated detections elsewhere in the same README.
• End-to-end, that means the benchmarked Rust repositories in this PR keep the intended repository semantics — for example Apache-2.0 OR MIT in atomic-rs, rust-base64, and the member crates inside aeads — instead of reporting the pair as a conjunction or surfacing extra singleton leftovers as the primary expression.

Follow-up work

• Created or intentionally deferred:
  • if additional benchmark targets surface the same kinds of reviewed differences, prefer scanner fixes or benchmark-note framing first; only introduce compare normalization when it is clearly generic and semantically lossless