Hull-edge seed evaluation (no-op verdict)

Evaluated seeding the active-set walk’s first-piece right-edge bisection (piece_window_right_edge_seeded, cl/crossings.rs) from the composed bound hull’s first growth-rate-takeover (bound slope crossing 1) to cut grow-loop probes.

Outcome: not shipped

The mechanism is sound (byte-identical goldens; the seed is an advisory hi feeding the same validated grow loop), but it misses production:

  • The walk’s event solver fully supersedes the seeded bisection at the production-default stance (event_solver_fallbacks = 0 on all measured corpora: heavy_cl_solve_captures, live_capture_loop13/17, synthetic cl_corpus).

  • Where the legacy stance runs, the measured win is small: −39/12 path probes per block (heavy corpus) against millions of probes spent elsewhere in the same corpora.

  • The hull’s first takeover frequently sits at the origin (surrogate peak at x = 0 for saturated paths), and where it exists it tends to overshoot the actual piece edge, so only a handful of probes move.

Maintenance cost (+174/−8 lines, four files) is not justified for a legacy-branch-only, low-double-digit probe saving. Revert, and keep this evaluation as the record for future ideas that would seed walk bisections from envelope geometry: measure under the production stance first.

Rejected landing: the seed computation now lives only in git history of the working tree that produced the measurements; no residual machinery.