Fix destination reachability and add shared-basis evaluation metrics

[sergio-gimenez]

Summary

Two coupled problems fixed, plus tooling for running the evaluation campaign:

1. Link-state could not reach deliverable ground stations. The Hypatia-derived fstate path collapsed per-GS visibility to the single nearest satellite before routing, which made the destination a fixed satellite rather than the ground station itself. Under sparse connectivity that satellite is frequently in an unreachable component while another visible one is reachable, so link-state reported failure on pairs that were deliverable — in a Ring topology it delivered 80 of 408 deliverable pairs instead of all of them.
2. The stretch baseline was algorithm-dependent. Each algorithm was graded against a shortest path to whichever egress satellite it happened to reach, over whichever subset of pairs it managed to deliver. Neither the subset nor its size was reported, so an algorithm that fails on the hard pairs was rewarded: only its successes reached the average.

Changes

Routing fix (6678dc6)

• Thread the full per-GS visibility list (ground_station_satellites_in_range) through algorithm_free_one_only_over_isls and the fstate calculation, so the destination egress is whichever visible satellite minimises path length plus GSL length.
• Single-attachment fallback retained for callers without a precomputed visibility list; it understates reachability and logs as such.
• Test call sites updated for the new parameter.

Metrics (f669f87)

• Reachability is now decided from the topology alone, before any algorithm is consulted, via _best_reachable_egress: the best end-to-end route to any satellite the destination can see (hop-optimal and distance-optimal egress tracked separately, since they need not coincide). Every algorithm is scored over the same pairs against the same lower bound.
• New delivery_* columns: deliverable pairs, delivered pairs, delivery rate, forwarding failures, and the separate causes of non-delivery (no_src_visibility, no_dst_visibility, disconnected).
• New stretch_hop_shared / stretch_dist_shared graded against the shared lower bound; legacy stretch_hop / stretch_dist kept for continuity with earlier runs.
• New delivery_non_optimal_egress_rate: share of delivered pairs exiting through a non-optimal egress — the quantity that separates the two stretch families.
• Semantics documented in docs/evaluation.md.

Tooling (03e7959)

• scripts/run-matrix-parallel.sh: config × algorithm × ISL-scenario matrix as parallel Docker jobs, JOBS-bounded, BLAS pinned to 1 thread per container, completed jobs skipped on rerun (interrupted campaigns resume), per-job wall-clock and exit code in job_timings.csv.

Verification

• tests/network_state/test_algorithm_free_one_only_over_isls.py: 4 passed.
• By construction, a shortest-path algorithm scores 1.0 on the shared basis — link-state doubles as a correctness check on the metric itself.

Chores

• dbbed6a backfills the CHANGELOG Unreleased section for the baseline-removal refactor (ad05c47, previously unrecorded) and this branch.