fix(mincut): report exact results and qualify release claims

[rudycelekli]

Summary

DynamicMinCut::recompute_min_cut always calls the exact sparse Stoer–Wagner solver. Its legacy .approximate(ε) option changed only MinCutResult.is_exact and approximation_ratio, so the same exact cut was mislabeled as approximate. This PR reports the algorithm actually run (is_exact: true, ratio 1.0) while retaining the builder option and configuration fields for compatibility. A regression compares values, partitions, and labels across exact and legacy settings after graph updates.

The public crate description, README, crate docs, and linked getting started/core concept guides now distinguish the exact polynomial recomputation path from separate research modules. Historical n^0.12 measurements and paper-component inventories are explicitly qualified; ApproxMinCut's ε-derived interval is described as heuristic rather than a certified all-graph bound. No benchmark result or paper implementation is claimed by this PR.

This addresses the misleading result labeling and public claim portions of #988. It does not close the issue: push-to-main release gates, independent benchmark receipts, and release qualification still need separate work. The fast-uri item was already addressed by #1014 per the maintainer's issue comment.

Reproduction and validation

• Before the fix, the new focused regression failed because .approximate(0.5) returned is_exact: false despite matching the exact solver's value and partition.
• After the fix, the focused algorithm::tests:: suite passed: 25 tests.
• Reduced standalone crate harness (same default exact/approximate features, avoiding unrelated sparse-workspace dependencies): 504 unit tests passed, 1 ignored, 13 subpolynomial:: tests filtered; 26 doctests passed, 20 ignored.
• rustfmt --check on changed Rust modules and git diff --check: passed.

The full workspace CI and long-running subpolynomial:: tests were not run locally. The reduced harness uses crates.io dependency resolution rather than the workspace lockfile; upstream CI remains the integration gate. No LocalKCut source from #1079 is changed.