Stacked on #1551 → #1552 → #1553 — this branch includes all three; its own content is the top 3 commits. First installment of the refactoring roadmap's Phase 2 (structural consolidation): three pilots, one per workstream, validating the patterns before the rest of the phase.

Summary

1. Context derived type (pilot for state-bundling). pre_process's five module-level initial-condition arrays (q_prim_vf, q_cons_vf, q_T_sf, bc_type, patch_id_fp) bundle into one ic_context in m_derived_types — plain data, no methods, mixed-precision variant preserved. Other modules' signatures are untouched (they already received these as arguments). +2 net LOC; emitted-Fortran equivalence verified (zero executable-statement diffs after qualifying the references).

2. Direction-unified BC dispatcher (pilot for de-triplication). s_populate_variables_buffers carried six near-identical x/y/z-beg/end blocks; they collapse into one parameterized s_populate_bc_direction(bc_dir, bc_loc, ...). −105 net LOC. Behavior preserved exactly per (direction, location) pair, including a subtlety the original encodes: only y-beg dispatches BC_AXIS and excludes it from qbmm extrapolation — y-end does not, and the unified guard reproduces that asymmetry. GPU directive census: six identical GPU_PARALLEL_LOOP shells become one (with one added private scalar); both backends, otherwise unchanged. The BC-code read is hoisted to a loop-private, and the bc_type edge is passed as an integer_field so the helper contains no constant direction subscripts (keeps case-optimized 1D builds compiling).

3. Module split (pilot for giant-file decomposition). m_boundary_common.fpp (2,250 lines) splits into the dispatcher+lifecycle core (194 lines), m_boundary_primitives (the eight GPU_ROUTINE BC helpers + the device-resident bc_buffers), and m_boundary_io (capillary/IGR buffers, MPI types, restart read/write, grid buffers). Pure code motion: statement-multiset of the union equals the parent file per target; GPU-directive multiset md5-identical. The original module re-exports everything it used to own, so zero external callers change. One deliberate deviation from the plan is documented in-code: bc_buffers lives in m_boundary_primitives (its on-device consumer) rather than the core module, avoiding a use-cycle with the re-export.

Verification

• Full golden suite (sharded, all ~590 cases) green on the complete stack.
• Per-pilot independent review with reproduced evidence: emitted-.f90 statement-set equivalence (pilots 1, 3), six-way per-(dir,loc) behavior table and GPU-directive census (pilot 2), re-export resolution audit covering all 13 external consumer files (pilot 3).
• Simulation grind times 0.99–1.00 vs parent across the benchmark suite (the dispatcher is hot-path; measured before the perf question was waived).

Notes for reviewers

• All three pilots are strictly behavior-preserving; no golden files changed.
• The patterns these pilots validate (context types for pre/post state, direction parameterization, seam-based file splits) are the templates for the remainder of Phase 2 (post-process contexts, m_riemann_solvers split) — feedback on the patterns here shapes those.