[bpy13]

Speeds up the solver about 30x (default run 2.6 s -> 0.1 s; all 14 paper runs 95 s -> 1.4 s) without changing the physics.

Changes

1. Faster Runge-Kutta step: same arithmetic, bit-for-bit identical results (1.36x).
2. Brent's method for the last stage of the central-density search (finding r_s = r_sb), to the same 1e-12 precision. rho_0 agrees with the paper code's bisection to 2e-12; profiles move by at most 2e-4 relative; the match to the published runs is 1.7e-4 (was 2.1e-4). find_central_density(..., method=bisection) keeps the paper code's search.
3. Figure tabs solve their runs in parallel processes when Numba is not installed (Fig. 7: 20-40 s -> 5.5 s). The app script is now guarded by if __name__ == __main__.
4. Optional Numba extra (pip install -e .[dev,fast]): a compiled copy of the trial integration in src/outflow/_compiled.py, bit-for-bit identical to the Python code on all 14 paper runs. With it, figure tabs solve their runs in-process (about 1.5 s).

Checks

• Local: 243 tests (fast and slow) and 27 legacy cross-checks pass; ruff and the docs check are clean.
• Compiled vs Python: identical on 45 inputs x 2 tolerances, and on all 14 full runs (slow tests); also on Linux in CI.
• Real app driven in headless Edge: Fig. 7 tab with and without the process pool, no errors.

Decisions and measurements: docs/decisions.md, section Speed.