TASK-052: stop advection transporting through solid walls

[AdamClemens]

Stacked on #91 (Stage 9's insertion) — merge that first.

FirstOrderUpwindAdvection took a genuine boundary face's transporting velocity from the cell inside it rather than from what the configuration prescribed there. GreenGaussDivergence meanwhile consulted the boundary condition and treated the same walls as solid — so the two operators disagreed about the domain they were solving in, and material crossed solid walls at up to 42% of the flow's own peak speed.

This falsified a use case Stage 4 wrote down for itself:

Prescribe a wall's value or its gradient, or make a boundary periodic, and have the interior schemes honour it.

The measurement

Shipped smoke_transport.yaml geometry, tracer diffusion zeroed so only advection remains. Pure advection in a closed domain must conserve exactly:

	before	after
Sealed box, 400 steps	−14.27%	+0.000000000000%

Evidence this is physics, not a re-fitted tolerance

The error against Ghia, Ghia & Shin (1982) fell at every resolution and stayed monotonic:

mesh	before	after
9×9	0.1433	0.1292
13×13	0.0874	0.0766
17×17	0.0578	0.0524

Both baselines are recorded rather than the old ones overwritten (Criterion 7). The 0.08 bound is deliberately unchanged — tightening it to each improvement would stop it being an independent check. Couette, Taylor-Green (both halves), kinetic-energy monotonicity and Stage 6's periodic domain integrals and decay rates were all confirmed unmoved rather than assumed so.

The fix is reuse, not new mechanism

boundary_normal_velocity is extracted from what divergence already did; advection now calls it too. Three files, no interface change, no registry change, no arity change anywhere.

A first implementation threaded a per-edge mapping from BoundaryFaceConfig.velocity through assemble_numerics and widened two registries. It was built, measured working, and abandoned when test_divergence_contract.py's linear-field exactness test turned out to be inexpressible through a per-edge scalar — a linear field's normal component varies along an edge, and a future parabolic inlet would have the same problem. That is recorded in the task entry rather than presented as the first idea.

A second defect, found by a test rather than by reading

DirichletBoundaryCondition.evaluate fell through to BoundaryFaceConfig.scalar_value for a vector field with no override — a transported scalar's boundary value, returned as a velocity. Every shipped config leaves it at 0.0, so walls were impermeable by coincidence of the defaults rather than by anything the configuration said: the same shape of accident the pre-fix divergence operator had. A new scenario failed against a fixture whose scalar boundary value was 3.0, because the wall then resolved to a normal velocity of 3.0.

One check moved, and it belonged to another stage

Stage 6's Rayleigh–Bénard onset scenario compared final vertical-velocity RMS and required a factor of two. It passed at 6.35 and now measures 1.28 — because that verdict was produced by this defect: heat and momentum were crossing that fixture's own solid walls with the Dirichlet condition re-injecting boundary-temperature fluid, a spurious heat pump.

Re-derived as a growth comparison (maintainer's call) — which is what that fixture's own docstring says it always meant to check, and which the 2026-08-31 audit found had never been implemented:

heated from below: 0.0904 → 0.1621   (+79%, still growing)
heated from above: 0.1251 → 0.1267   (+1.3%, settled)

Roughly sixty to one, against bars of 25% and 5%. Mutation-verified: inverting the buoyancy sign drops "below" to 1.012 and fails it. Recorded in Stage 6's own status table as a verdict a later stage changed, not rewritten as though it had always read that way.

Honest about what is not done

Criterion 3 is discharged only partially, and the Discharges section says so. The sweep over BoundaryFaceConfig's own fields is not built, and BoundaryFaceConfig.velocity still reaches no scheme — it is validated for mutual exclusivity and zero net flux, then read by nothing. What this task closes is the half that was actually wrong.

TASK-053 (a failed frame failing the run) and TASK-054 (the timestep stability warning) are drafted, not built — both carry Status: Not started, so Stage 9 reads as opened rather than complete.

Verification

make ci green in full: 1206 passed, 20 skipped, 156 scenarios, all 15 structural checks. All 11 golden demos re-run headlessly; sealed_box verified through pyflow run --demos sealed_box and present in the demo listing as #12; the documented record → resume → play pipeline re-run by hand.

🤖 Generated with Claude Code