Unify per-solver elastic interface physics after the formulation rulings

[sbryngelson]

Motivation

Four findings from the 2026-06 series share one root cause: elastic feature physics was implemented per-Riemann-solver and the copies drifted independently —

• Hyperelastic wave-speed estimates: HLLC includes the tau_e correction, HLL deliberately omits it #1557: HLLC applies the tau_e wave-speed correction to hyperelastic runs; HLL deliberately omits it (live divergence).
• HLL hypoelastic damage load reads the LEFT cell's damage for the RIGHT state (missing +1 offset) #1570: the damage scaling reads the left cell's damage for the right state (missing offset) — possibly intentional, contradicts every neighboring load.
• HLL hyperelastic wave-speed branch reads shear moduli that only the hypoelastic path initializes #1571: HLL's hyperelastic wave-speed branch reads shear moduli only the hypoelastic path initializes (live uninitialized read).
• HLLC hypoelastic energy block mis-doubles the theta-theta normal stress in 2D-axisymmetric (currently dead code) #1566 (fixed by deletion in Phase-3: Riemann hot-path decomposition into shared GPU device helpers #1572): HLLC's dead hypoelastic block carried a 2D-axisymmetric shear-doubling bug.

#1572 unified the hypoelastic interface energy under maintainer rulings and proved the pattern: one GPU_ROUTINE helper, dead copies deleted, golden suite + <5% benchmark gates. The wave-speed and flux blocks remain per-solver, and the open questions above block unifying them — these are physics formulation decisions, not refactoring decisions.

Proposal

After the elasticity authors rule on #1557/#1570/#1571:

1. Unify the elastic wave-speed estimation into a shared helper in m_riemann_state (the ruled-correct formulation), parameterized only where the ruling says solvers should genuinely differ.
2. Delete the remaining dead hypoelastic flux/stress blocks in LF and HLLC (the validator restricts hypoelasticity to HLL; the blocks read state nothing writes since Phase-3: Riemann hot-path decomposition into shared GPU device helpers #1572's deletions) — net-negative LOC, zero behavior.
3. Add the Fortran-side backstop: an @:PROHIBIT mirroring the hypoelasticity-requires-HLL restriction in src/simulation/m_checker.fpp, so the toolchain-bypass route (HLL hyperelastic wave-speed branch reads shear moduli that only the hypoelastic path initializes #1571's reachability) is closed regardless of the constraint-unification proposal's timeline.
4. Apply the offset fix from the HLL hypoelastic damage load reads the LEFT cell's damage for the RIGHT state (missing +1 offset) #1570 ruling (a one-character change if unintentional).

Execution sketch

Identical machinery to #1572: investigation-first mapping of every elastic block per solver, rulings encoded as binding constraints in the implementation plan, helper extraction with the caller-loads/helper-computes pattern (documented in gpuParallelization.md), per-commit emitted-equivalence where behavior-preserving and explicit behavior notes where the rulings change results, golden suite + per-solver benchmark gates (the hypo_hll case directly exercises this path).

Effort: small-moderate once unblocked. Risk: low for items 2-3 (deletion + checker); item 1's risk is owned by the physics ruling, not the refactor. Blocked on: #1557, #1570, #1571.