diff --git a/docs/refactoring_plan.md b/docs/refactoring_plan.md
new file mode 100644
index 0000000..8f495fe
--- /dev/null
+++ b/docs/refactoring_plan.md
@@ -0,0 +1,102 @@
+# CppNCorr Refactoring Plan — Round 2 (`ja/refactoring/2`)
+
+Status: **proposal** — this document is the working plan for the second refactoring
+round. It supersedes the stale `ja/refactoring/1` branch (PR #10), which forked
+before the newversion restructure (#12) and no longer applies to `main`.
+
+## 1. Where round 1 stands
+
+`ja/refactoring/1` (9 commits, forked at `74a6deb`) pursued two themes:
+
+1. **Extraction by algorithm** — seed analysis + serialization + visualization
+   split out of `ncorr.cpp` (5815 → 3386 lines), `ImageProcessor` /
+   `VideoImporter` modules, opt-in diagnostics (`NCORR_ENABLE_DIAGNOSTICS`).
+2. **Extraction by dependency** — `Array2D.h` split into OpenCV / FFTW / BLAS /
+   linear-solver headers, plus an umbrella `include/ncorr/` forwarding-header tree.
+
+Since then `main` absorbed the newversion (#12: in-memory API, params, tests,
+docs, CI), the exact quintic interpolation / fake-jump fix, the recursive
+B-spline filter, multi-reference chaining, and leveled logging (#18). Verdict
+per theme:
+
+| Round-1 theme | Verdict |
+|---|---|
+| ImageProcessor / VideoImporter / FilterConfig | **Already integrated** — now static members / factories in `Image2D.h` (in-memory modes, `from_mat()`, `from_file_filtered()`) |
+| Diagnostics env-var system | **Superseded** — `ncorr/log.h` leveled logging (#18) is strictly better |
+| Seed analysis module | **Still open, high value** — must be redone against main (both `matlab_*` and `exact_matlab_*` variants now exist) |
+| Serialization module | **Still open, moderate value** |
+| Array2D dependency split | **Defer** — header-only templates; split buys no compile-time or link benefit today; revisit only if cuNCorr needs a subset |
+| Umbrella `ncorr/core/*.hpp` forwarding headers | **Drop** — main's pragmatic layout (flat legacy headers + `include/ncorr/` for new modules) is fine |
+
+The branch diff no longer applies (`ncorr.cpp` was rewritten underneath it);
+nothing is worth cherry-picking as code. The ideas below are re-implemented
+against current `main`.
+
+## 2. Current pain points on `main`
+
+- `src/ncorr.cpp` — **5815 lines**: DIC engine (~1258 ln), I/O + visualization
+  (~650 ln), seed-based parallel DIC (~340 ln), chain composition
+  (~790 ln: `add_with_rois` + `exact_*` + legacy), helpers/optimizers (~2780 ln).
+- `include/Array2D.h` — **4241 lines**, monolithic but well-sectioned.
+- **Duplication**: `matlab_DIC_analysis_impl()` (ncorr.cpp:5085) vs
+  `exact_matlab_DIC_analysis_impl()` (ncorr.cpp:5684) share ~65% of their
+  bodies; they differ only in the chain-composition call
+  (`add_with_rois` vs `exact_add_with_rois`).
+- **Missing test seams**: seed scheduling and chain composition are private
+  helpers, only covered end-to-end (`test/src/ncorr_test_chain_seam.cpp`).
+- **Unclear API surface**: `DIC_analysis()`, `matlab_DIC_analysis_*()` and
+  `exact_matlab_DIC_analysis_*()` are all public; optimizers
+  (`subregion_nloptimizer`, `disp_nloptimizer`) live in `ncorr.h` without a
+  documented contract. This matters now that CPPxDIC treats CppNCorr as one of
+  two interchangeable DIC engines (alongside cuNCorr).
+
+## 3. Plan
+
+### Phase 1 — extract by algorithm (high value)
+
+1. **Seed analysis module** (`include/ncorr_seed.hpp` + `src/ncorr_seed.cpp`)
+   - Move `matlab_compute_seed_segment()` (ncorr.cpp:4877), `matlab_run_segment_dic()`
+     (ncorr.cpp:4997), the `matlab_seed_*` helpers (ncorr.cpp:4804–4875) and the
+     `matlab_seed_segment` struct behind a small API
+     (`compute_seed_segment()`, `run_segment_dic()`).
+   - Must serve both the canonical and `exact_*` analysis paths.
+   - Add unit tests for seed scheduling (currently untestable in isolation).
+   - Est. ~300 lines moved; ncorr.cpp shrinks accordingly.
+
+2. **Exact chain composition module** (`include/ncorr_exact_chain.hpp` + `.cpp`)
+   - Move the `exact_detail` namespace (ncorr.cpp:5409–5551),
+     `exact_add_with_rois()` (ncorr.cpp:5556) and
+     `exact_matlab_DIC_analysis_*` (ncorr.cpp:5684–5815) into a dedicated
+     `ncorr::exact` module. Makes explicit that the MATLAB byte-for-byte path
+     is an *alternative strategy*, not part of the core engine.
+   - Existing regression test (`ncorr_test_chain_seam.cpp`) guards the move.
+   - Est. ~400 lines moved.
+
+### Phase 2 — de-duplicate (clarity)
+
+3. **Unify `matlab_DIC_analysis` variants** via a chain-composition strategy
+   parameter (template or callable): one `matlab_DIC_analysis_impl(input,
+   parallel, chain_strategy)` with canonical and exact strategies. Removes the
+   ~65% duplicated body; the substitution point becomes explicit.
+
+4. **Document/tighten the public API** — clarify in `ncorr.h` + developer guide
+   which entry points are the supported engine surface (what CPPxDIC's adapter
+   calls) vs internal; move internal optimizer declarations out of the public
+   header if feasible.
+
+### Phase 3 — optional / deferred
+
+5. **I/O module extraction** (`DIC_analysis_step_data` save/load,
+   Disp2D/ROI2D/Strain2D stream operators, ~600 lines) — only if format
+   versioning or isolated I/O tests become necessary.
+6. **Array2D split** — deferred; revisit if the cuNCorr port needs a
+   CPU-reference subset of Array2D without FFTW/OpenCV.
+
+### Guardrails
+
+- One extraction per PR-sized commit; each keeps the full test suite green.
+- No behavior changes: extractions are pure moves + renames; the strategy
+  unification (item 3) must be provably equivalent (existing regression tests
+  + chain-seam test).
+- Public header layout stays flat for legacy types; new modules go under
+  `include/ncorr/` per the newversion convention.