Rotated workspace cache follow-ups from the #543 review: config latch into the cached KSP (m1), enumeration reuse by coincidence (m3)

[lmoresi]

Two non-blocking findings from the #543 adversarial review of the rotated-workspace reuse (full detail in the review comment there):

1. m1 — solver.tolerance and a custom-FMG registration latch into the cached KSP at ctx creation. A later tightening is still honoured on the true residual (measured 1e-4 → 1e-10, err 1.8e-11) but via extra outer increments, with the inner-margin discipline silently violated. The cache key (or the ctx) should observe the tolerance/FMG configuration and rebuild the KSP when it changes.
2. m3 — the coefficient-variable enumeration is reused on geometry match alone; today this is safe only because every dependency-changing route forces a full _build teardown — a coincidence inherited from aux-layout invalidation, not enforced by the cache. Re-derive the enumeration on JIT-key mismatch so the safety is by construction.

Underworld development team with AI support from Claude Code