[cekees]

C5 steps 1 and 2 (claim A1, PARALLEL_SESSIONS §10), following YMF_MODEL_FAMILIES_C42 §6 D1 (dispatch) and D8 (hints are checked, and dispatch wins).

A ymf spec states the model, the formulation (a weak_forms entry named by from_weak_form), and the discretization: spaces, time integration, stabilization, sub-problems, bound symbols and regularization. proteus.Dispatch reads one discretization of a closed spec. For each sub-problem, in solve order, it returns the mprans module, the Coefficients flags, and the physics- and numerics-module attributes. Each value records where it came from: cell, spec, default or hint. The table, proteus/DispatchTable.py, is data only and can be reviewed on its own.

commit	what	measured
1bf365b	DispatchTable.py (13 cells: RANS2P ×2 forms, RANS3PF + PresInc/Pres/PresInit, VOF, NCLS, RDLS, MCorr, CLSVOF, SW2DCV, GN_SW2DCV), Dispatch.py, test/dispatch	20 tests pass. Object-by-object match with the decks' built p/n modules: level_set_five = damBreakLevelSet (5 models, 100 attributes), level_set_clsvof = damBreakLevelSet interface=1 (2, 49), clsvof_projection = damBreak (5, 75), SW2DCV = dam3Bumps (23), GN_SW2DCV = dSWEs_steady_state (16). 10 of 10 single-value mutations are caught.
267de2b	ymf_run refuses split discretizations before adr_problem and prints what each one dispatches to	before: a SymbolicError traceback on the dam break and the lake at rest; after: exit 2 with the dispatch printed. test_pro_cli + test_adr_problem + test_domain_ymf + test/dispatch: 213 passed, 1 skipped
68f5df0	.github/coverage/floors.json: NCLS3P.py 76.2 → 75.6	Not this PR's code. 76f3e12 on main removed 15 covered statements from NCLS3P (685 → 670) and the same 163 lines are still missed, so 76.2% → 75.67%. The floor is set to floor(10·now)/10. Every branch off main c09280c hits this check.
6f0d07c	Step 2, two-phase: ymf_run runs a split discretization from its dispatch. proteus/DispatchRun.py builds a TwoPhaseFlowProblem from the dispatch and from ymf's split_problem: Tank2D + BC_RANS setters matched on the closed conditions, ICs (Proteus's smoothedHeaviside for heaviside(φ)), and ρ/ν/g from the family's ρ, μ, g. It checks the built p/n against the dispatch before solving; the archive records the dispatch and the realization, and --check works	dambreak_short.ymf (level_set_five, 3 outputs, h = 0.2576) vs damBreakLevelSet.py -C "final_time=0.03 dt_output=0.01 he=0.08 genMesh=True": 103/103 datasets bitwise equal on 1 rank, 141/141 on 2 (-mat_partitioning_type current). --check reproduces bitwise. Negative control ε = 1.6 h: vof_t0 differs by 1.7e-2. Needs ymf's hypot (below)
f3f0078	Step 2, shallow water: an SWFlowProblem with SWFlow's own sw_p/sw_n loaded through Context; walls → reflectingBCs. A cells-meshed split discretization is refused up front	dam3bumps.ymf (t ∈ (0, 0.5], 5 outputs, h = 2, 969 nodes) vs dam3Bumps.py -C "final_time=0.5 dt_output=0.1 structured=False he=2.0": 70/70 numeric datasets bitwise equal on 1 rank, 77/77 on 2. For scale, the deck on 1 vs 2 ranks differs by up to 4.3e-6. Runs on ymf main
1cbe30c	ymf_run refuses, before anything runs, a split discretization the realizer would refuse on the data alone (DispatchRun.precheck: family, box, boundary setters, two-phase values)	level_set_clsvof: before, a traceback mid-solve (exit 1); after, exit 2 with nothing written. The reason: no sub-problem carries χ, so the closed atmosphere has no χ inflow, and no BC_RANS setter matches {p = 0, zero velocity flux}. 27 passed (hypot) / 24 + 3 skipped (ymf main)
bed1cef	test/dispatch: no skip without ymf's hypot; the dam-break oracle needs cekees/ymf#12	ymf notation-hypot (6feb177): 27 passed. ymf main (c2e646f): 24 passed, 3 failed (the dam-break runs on 1 and 2 ranks, and the ymf_run test)

test_ymf_run_says_what_a_split_discretization_dispatches_to is in commit 1's test file and needs commit 2. Since 6f0d07c it describes the run: what the split dispatched to and what the archive records. The refusal has its own tests: an uncovered cell, a cells-meshed split, and a boundary no setter imposes.

Needs cekees/ymf#12 merged into proteustoolkit/ymf main first. It adds hypot(a, b) to ymf's notation. damBreakLevelSet.py's phi_init uses hypot; the spec's sqrt-of-squares form differs from it by 1 ulp (4.4e-16) at 2 of the 109 stored-mesh nodes, and by t = 0.03 that grows to 1182 values, up to 1.5e-12 in p. With hypot, 0 values differ. proteus CI installs ymf from proteustoolkit/ymf@main, so until ymf#12 merges upstream the dam-break runs and the in-process ymf_run test cannot pass on CI. They are not skipped: until then they fail on CI (unknown name 'hypot'), which is the intended order. The dam3Bumps runs do not need it.

Finding: both in-tree dam breaks solve the non-conservative inertia form. RANS2P.Coefficients defaults to NONCONSERVATIVE_FORM=1.0, and RANS3PF.h hardwires that form. The C42 note's from_weak_form: conservative_inertia ("to confirm") is therefore wrong for both. (conservative, P2 projection) is a cell no module covers, and dispatch refuses it by name.

How a sub-problem finds its cell: the family comes from physical_model. The role comes from the closed spec's structure: which unknowns its equations are for, and what it enforces or introduces. The cell is then chosen by formulation and by each option the spec states. Values the dispatcher infers from typed keys:

• epsFact and the other widths, from regularize widths in element sizes
• useVF, from lagging χ = heaviside(φ)
• shock capturing, from stabilization
• maxNonlinearIts, from solver
• quadrature, from the highest space order

Refusals tested:

• the conservative projection
• a P2 RANS2P velocity
• SSP_RK3 on VOF
• a model with no family, a sub-problem with no role, an unknown formulation
• RANS3PF without PresInit
• an unregularized heaviside
• an unread SWE scheme option
• hints that disagree with the spec (useVF, RDLS epsFact) or with the cell (NONCONSERVATIVE_FORM), name another module, or are unknown to the cell

Not in this PR: the CLSVOF and projection discretizations as runs. They dispatch, and the realizer should build them, but no oracle has been run for them yet. Also not here: the dispersive SWE (refused, since nothing realizes it), gauges, open SWE boundaries, 3-D tanks, and stored meshes as a spec domain (C7.1).

Run locally: env proteus-bc-261025, macOS arm64; ymf main c2e646f, and ymf notation-hypot d51ca2d for the dam-break runs.

🤖 Generated with Claude Code