Make PISO's pressure correction genuinely multi-pass (TASK-033)

[AdamClemens]

Summary

• Stage 5's fourth task: turns TASK-027's single-pass PISO into a real corrector loop. Each pass records its own maximum divergence; the loop returns once it's at or below numerics.pressure_correction_tolerance, or raises DivergenceDidNotConvergeError after numerics.pressure_correction_max_iterations rather than returning a best-effort result.
• Resolves Stage 5's design question three (what carries the momentum-equation coefficients Rhie-Chow needs, given PyFlow's fully explicit RK4 predictor) by numerical prototyping before writing any test or implementation code, per the roadmap's own instruction: a_P = V/dt is the only contribution to a_P for this architecture, and PyFlow's uniform cell volume makes V/a_P = dt one constant for the whole mesh. Pairing that correction with the same compact Laplacian the Poisson matrix already uses — not the composed Gradient/Divergence pair TASK-027 tried and measured stalling — restores the discrete adjoint property and converges a manufactured field to floating-point-exact zero divergence in one pass, verified numerically before being trusted.
• tolerance/max_iterations are bound at PISO's own construction ("outer-loop state the strategy owns"), so PressureCoupling.correct's interface is unchanged and no ADR was needed.
• Couette flow's own comparison stays with TASK-034 alone, which has the fully assembled timestep it needs — this task supplies the corrector loop it depends on.

Test plan

• tests/features/pressure_correction_loop.feature written first (TDD), three scenarios: monotonic convergence with a real solver, genuine multi-pass convergence with a deliberately halving solver, honest exhaustion (DivergenceDidNotConvergeError) with a no-op solver.
• tests/unit/test_pressure_correction_loop.py binds all three scenarios against the real PISO.
• New NumericsConfig.pressure_correction_tolerance/pressure_correction_max_iterations fields, with load/reject tests.
• Blast radius: assembly.py's register_pressure_coupling widened to a four-argument factory (own stale docstrings found and fixed in the same pass); _StubLinearSolver/_CapturingPressureCoupling test doubles updated; docs/architecture/icds.md, docs/architecture/engine.md, docs/repository-manifest.md, and all relevant CLAUDE.md files updated to describe PISO as genuinely multi-pass rather than single-pass.
• make ci green: 653 tests passed, 99% coverage, all doc/graph/scenario/status checks pass.

🤖 Generated with Claude Code