Classes
class	TestAccelMatchesPython
class	TestRandomizedSweep
class	TestCoolingIsActuallyApplied
class	TestDeliveredIspInvariant
class	TestChugKernelMatchesPython

Functions
	_chamber_gate ()
	_python_only ()
	_rel (got, want)
	_assert_close (name, got, want, rtol=RTOL)
	_py_field (ref, key)
	_rig (cfg_rel)
	_default_path (r, p)
	_need_numba ()
	test_injector_solve_matches_python (cfg_rel)
	test_default_path_matches_python_on_the_shipped_engine ()
	test_chamber_kernel_serves_the_designs ()
	test_heatup_tables_match_coolprop ()

Variables
	ROOT = Path(__file__).resolve().parents[1]
float	PSI_TO_PA = 6894.76
float	PA_AMBIENT = 101325.0
	_ab_parity
list	POINTS_PSI = [(563.467, 567.644), (518.4, 550.6), (597.3, 584.7)]
int	RTOL = 1e-6
list	CONFIGS
list	CORE_FIELDS
list	DIAG_FIELDS
dict	_RIGS = {}
list	_INJ_CONFIGS
int	_INJ_CORE_KEYS = 18
list	_INJ_KEYS

Detailed Description

Live A/B parity for the Numba accelerator against the authoritative Python physics.

Runs both implementations on the same inputs at test time and diffs them field by
field, so a drift on EITHER side fails regardless of which one moved. This is the
sole numeric guard on the accelerated path now that the C port is gone.

TOLERANCE. RTOL is 1e-6, not the 2e-3 the retired C suite used. Measured agreement
on 120 randomized points across both configs is ~2.5e-9 -- the accelerated and
Python paths run the same physics and differ only by Brent's convergence
tolerance, so 1e-6 leaves ~400x headroom while still being three orders tighter
than the old contract. If a change pushes this above 1e-6, that is a physics
divergence, not rounding: fix it rather than widening the bound.

THE REFERENCE MUST BE FORCED TO PYTHON. With the accelerator enabled,
runner.evaluate reaches chamber_solver._accel_chamber_pc -> accel.chamber_solve
and closure.flows -> accel.solve, so an unguarded "Python" reference is largely
the same numba kernels and the comparison is self-referential (it reads as ~1e-15
agreement, which is the tell). _python_only() below disables the accelerator for
the reference computation; without it this suite proves nothing.

THE CONFIG LIST IS DELIBERATE. canonical/impinging.yaml has ablative cooling ON and
a back-channel ring manifold, impinging_lox_ch4_8000N.yaml has neither (plenum, no
liner), canonical/pintle.yaml exercises the pintle injector -- a different solve
(kernels.injector_solve_pintle) and a different mixing input: no Rupe M, so E_m is
rupe_Em_opt flat and the drops leave axially -- and the 6.5 kN ethalox fixture is the
design as drawn, where both liquids heat up on CoolProp tables (engine.accel.chamber).

Function Documentation

◆ _assert_close()

test_numba_ab_parity._assert_close	(	name,
		got,
		want,
		rtol = RTOL )

protected

Definition at line 104 of file test_numba_ab_parity.py.

References _rel().

Referenced by test_default_path_matches_python_on_the_shipped_engine(), test_numba_ab_parity.TestAccelMatchesPython.test_diagnostics(), test_injector_solve_matches_python(), test_numba_ab_parity.TestAccelMatchesPython.test_kernel_level_raw_tuple(), and test_numba_ab_parity.TestAccelMatchesPython.test_wrapper_fields().

◆ _chamber_gate()

test_numba_ab_parity._chamber_gate ( )

protected

Python chamber physics the kernels do not mirror yet (empty => chamber kernels in use).

Definition at line 52 of file test_numba_ab_parity.py.

Referenced by _default_path(), _rig(), test_numba_ab_parity.TestCoolingIsActuallyApplied.test_effective_tc_differs_and_is_reported(), test_numba_ab_parity.TestAccelMatchesPython.test_kernel_level_raw_tuple(), test_numba_ab_parity.TestRandomizedSweep.test_sweep(), and test_numba_ab_parity.TestAccelMatchesPython.test_wrapper_fields().

◆ _default_path()

test_numba_ab_parity._default_path	(		r,
			p )

protected

What Layer 1 actually consumes: accel.evaluate, else (None) the runner on the DEFAULT,
accelerator-enabled path -- which still routes the injector through accel.solve and the chug
scan through the kernel.

Definition at line 148 of file test_numba_ab_parity.py.

References _chamber_gate(), and split().

Referenced by test_numba_ab_parity.TestAccelMatchesPython.test_diagnostics(), test_numba_ab_parity.TestRandomizedSweep.test_sweep(), and test_numba_ab_parity.TestAccelMatchesPython.test_wrapper_fields().

◆ _need_numba()

test_numba_ab_parity._need_numba ( )

protected

Definition at line 330 of file test_numba_ab_parity.py.

Referenced by test_chamber_kernel_serves_the_designs(), test_default_path_matches_python_on_the_shipped_engine(), test_heatup_tables_match_coolprop(), test_injector_solve_matches_python(), test_numba_ab_parity.TestChugKernelMatchesPython.test_minus_3pi_crossing(), and test_numba_ab_parity.TestChugKernelMatchesPython.test_randomised_loops().

◆ _py_field()

test_numba_ab_parity._py_field	(		ref,
			key )

protected

Definition at line 110 of file test_numba_ab_parity.py.

Referenced by test_default_path_matches_python_on_the_shipped_engine(), test_numba_ab_parity.TestAccelMatchesPython.test_kernel_level_raw_tuple(), test_numba_ab_parity.TestRandomizedSweep.test_sweep(), and test_numba_ab_parity.TestAccelMatchesPython.test_wrapper_fields().

◆ _python_only()

test_numba_ab_parity._python_only ( )

protected

Force the authoritative Python path for the reference computation.

Disables strict mode as well as the accelerator. Under ED_REQUIRE_ACCEL=1 (the
CI parity job) closure._try_native_flows treats "disabled + strict" as a
genuine accelerator failure and raises -- which is exactly right in
production, and exactly wrong here, where the accelerator is off ON PURPOSE.
Without this the whole suite fails under CI's env while passing locally.

Definition at line 81 of file test_numba_ab_parity.py.

Referenced by _rig(), test_default_path_matches_python_on_the_shipped_engine(), and test_numba_ab_parity.TestRandomizedSweep.test_sweep().

◆ _rel()

test_numba_ab_parity._rel	(		got,
			want )

protected

Definition at line 100 of file test_numba_ab_parity.py.

References max.

Referenced by _assert_close(), test_numba_ab_parity.TestChugKernelMatchesPython._same(), test_numba_ab_parity.TestCoolingIsActuallyApplied.test_effective_tc_differs_and_is_reported(), and test_numba_ab_parity.TestRandomizedSweep.test_sweep().

◆ _rig()

test_numba_ab_parity._rig ( cfg_rel )

protected

Config + runner + per-point pure-Python reference, built once.

Definition at line 119 of file test_numba_ab_parity.py.

References _chamber_gate(), and _python_only().

Referenced by test_numba_ab_parity.TestDeliveredIspInvariant.test_delivered_below_ceiling(), test_numba_ab_parity.TestAccelMatchesPython.test_diagnostics(), test_numba_ab_parity.TestCoolingIsActuallyApplied.test_effective_tc_differs_and_is_reported(), test_numba_ab_parity.TestAccelMatchesPython.test_kernel_level_raw_tuple(), test_numba_ab_parity.TestRandomizedSweep.test_sweep(), and test_numba_ab_parity.TestAccelMatchesPython.test_wrapper_fields().

◆ test_chamber_kernel_serves_the_designs()

test_numba_ab_parity.test_chamber_kernel_serves_the_designs ( )

Every chamber piece is mirrored, so the shipped design and the one drawn from CAD run on
the kernel (NOT_HANDLED here would put Layer 1 back on the ~60 ms Python solve).

Definition at line 470 of file test_numba_ab_parity.py.

References _need_numba().

◆ test_default_path_matches_python_on_the_shipped_engine()

test_numba_ab_parity.test_default_path_matches_python_on_the_shipped_engine ( )

The shipped 6.5 kN at its tank pressure and pad ambient: what the default (accelerated)
path returns -- the thing Layer 1 and the UI read -- must be the ED_ACCEL=off answer.

Definition at line 444 of file test_numba_ab_parity.py.

References _assert_close(), _need_numba(), _py_field(), and _python_only().

◆ test_heatup_tables_match_coolprop()

test_numba_ab_parity.test_heatup_tables_match_coolprop ( )

The one tabulation in the chamber kernel: CoolProp's liquid cp(T) and density at the
chamber pressure (combustion_physics.liquid_heatup_properties), for both liquids of the 6.5 kN
engine over the whole Pc window. A point the table refuses is allowed (the kernel returns no
solution and the caller runs Python); a point it answers must answer right.

Definition at line 486 of file test_numba_ab_parity.py.

References _need_numba().

◆ test_injector_solve_matches_python()

test_numba_ab_parity.test_injector_solve_matches_python ( cfg_rel )

closure.flows runs accel.solve on the DEFAULT path, on every residual of the Python chamber
solve. It must reproduce the Python injector: one closure pass with Cd untouched (impinging),
the feed exit dump K_exit on d_exit, the passage at A_hydraulic, pintle's fixed-K x*.

Definition at line 409 of file test_numba_ab_parity.py.

References _assert_close(), and _need_numba().

Variable Documentation

◆ _ab_parity

test_numba_ab_parity._ab_parity

protected

Initial value:

=  pytest.mark.skipif(
    os.environ.get("ED_REQUIRE_ACCEL") != "1" and os.environ.get("ED_AB_PARITY") != "1",
    reason="A/B parity runs in the accel-parity CI job; set ED_AB_PARITY=1 to run locally",
)

Definition at line 46 of file test_numba_ab_parity.py.

◆ _INJ_CONFIGS

list test_numba_ab_parity._INJ_CONFIGS

protected

Initial value:

=  ["configs/ethalox_6500N.yaml", "configs/canonical/impinging.yaml",
                "configs/impinging_lox_ch4_8000N.yaml", "configs/canonical/pintle.yaml",
                "tests/fixtures/ethalox_6500N_doublet_cad_2026-09-28.yaml"]

Definition at line 388 of file test_numba_ab_parity.py.

◆ _INJ_CORE_KEYS

int test_numba_ab_parity._INJ_CORE_KEYS = 18

protected

Definition at line 391 of file test_numba_ab_parity.py.

◆ _INJ_KEYS

list test_numba_ab_parity._INJ_KEYS

protected

Initial value:

=  ["Cd_O", "Cd_F", "delta_p_feed_O", "delta_p_feed_F", "delta_p_injector_O",
             "delta_p_injector_F", "D32_O", "D32_F", "x_star", "u_O", "u_F", "We_O", "We_F",
             "momentum_ratio_R", "rupe_M", "u_axial_spray", "iterations",
             "constraints_satisfied",
             # spray/doublet geometry Layer 1 reads out of the diagnostics (geometry-fit term,
             # final report); None on the accelerated path until the extras were added
             "L_imp", "D_pitch_O", "D_pitch_F", "element_gap_O", "element_gap_F", "s_pair",
             "vaporization_length_total", "L_sheet_breakup", "k_evap_O", "k_evap_F",
             "tau_evap_O", "tau_evap_F", "J", "TMR", "theta", "Oh_O", "Oh_F", "rho_gas_breakup",
             # the back-channel ring, hole by hole (impinging._manifold_diagnostics)
             "element_flow_ratio_min_O", "element_flow_ratio_max_O", "element_flow_ratio_min_F",
             "element_flow_ratio_max_F", "element_mixture_ratio_min", "element_mixture_ratio_max",
             "manifold_branch_velocity_O", "manifold_branch_velocity_F", "Cd_eff_manifold_O",
             "Cd_eff_manifold_F"]

Definition at line 392 of file test_numba_ab_parity.py.

◆ _RIGS

dict test_numba_ab_parity._RIGS = {}

protected

Definition at line 116 of file test_numba_ab_parity.py.

◆ CONFIGS

list test_numba_ab_parity.CONFIGS

Initial value:

=  [
    ("configs/canonical/impinging.yaml", True),          # impinging, ring manifold, ablative ON
    ("configs/impinging_lox_ch4_8000N.yaml", False),     # impinging, plenum, ablative off
    ("configs/canonical/pintle.yaml", True),             # pintle, ablative ON
    # the 6.5 kN LOX/ethanol doublet as drawn (ring manifold, CoolProp heat-up of both liquids)
    ("tests/fixtures/ethalox_6500N_doublet_cad_2026-09-28.yaml", True),
]

Definition at line 61 of file test_numba_ab_parity.py.

◆ CORE_FIELDS

list test_numba_ab_parity.CORE_FIELDS

Initial value:

=  ["Pc", "F", "Isp", "MR", "cstar_actual", "eta_cstar",
               "mdot_total", "mdot_O", "mdot_F", "Cf_actual",
               "P_exit", "T_exit", "v_exit"]

Definition at line 69 of file test_numba_ab_parity.py.

◆ DIAG_FIELDS

list test_numba_ab_parity.DIAG_FIELDS

Initial value:

=  ["D32_O", "D32_F", "Cd_O", "Cd_F", "momentum_ratio_R",
               "delta_p_feed_O", "delta_p_feed_F", "delta_p_injector_O",
               "delta_p_injector_F", "A_geom_O", "A_geom_F", "A_eff_O", "A_eff_F",
               "turbulence_intensity_mix", "u_O", "u_F", "We_O", "We_F"]

Definition at line 74 of file test_numba_ab_parity.py.

◆ PA_AMBIENT

float test_numba_ab_parity.PA_AMBIENT = 101325.0

Definition at line 39 of file test_numba_ab_parity.py.

◆ POINTS_PSI

list test_numba_ab_parity.POINTS_PSI = [(563.467, 567.644), (518.4, 550.6), (597.3, 584.7)]

Definition at line 57 of file test_numba_ab_parity.py.

◆ PSI_TO_PA

float test_numba_ab_parity.PSI_TO_PA = 6894.76

Definition at line 38 of file test_numba_ab_parity.py.

◆ ROOT

test_numba_ab_parity.ROOT = Path(__file__).resolve().parents[1]

Definition at line 37 of file test_numba_ab_parity.py.

◆ RTOL

int test_numba_ab_parity.RTOL = 1e-6

Definition at line 59 of file test_numba_ab_parity.py.

Classes

Functions

Variables

Detailed Description

Function Documentation

◆ _assert_close()

◆ _chamber_gate()

◆ _default_path()

◆ _need_numba()

◆ _py_field()

◆ _python_only()

◆ _rel()

◆ _rig()

◆ test_chamber_kernel_serves_the_designs()

◆ test_default_path_matches_python_on_the_shipped_engine()

◆ test_heatup_tables_match_coolprop()

◆ test_injector_solve_matches_python()

Variable Documentation

◆ _ab_parity

◆ _INJ_CONFIGS

◆ _INJ_CORE_KEYS

◆ _INJ_KEYS

◆ _RIGS

◆ CONFIGS

◆ CORE_FIELDS

◆ DIAG_FIELDS

◆ PA_AMBIENT

◆ POINTS_PSI

◆ PSI_TO_PA

◆ ROOT

◆ RTOL