PERRY_L14_NBC_ORDER (landed in merge train 254, 2dda7c1139 / #10929) is default-OFF, and shape_numeric_inputs() early-returns empty unless it is set:
// crates/perry-codegen/src/collectors/number_by_construction.rs:796
if !nbc_order_enabled() || shape_proven.is_empty() { return ...; }grep -rn PERRY_L14_NBC_ORDER crates/ finds exactly four sites — the gate function, its doc comment, and the two cache registries (build_cache.rs, object_cache.rs). No test sets the variable.
So the reorder the PR exists to perform is executed by no suite, and neither is the rule most likely to be wrong — that the shape inputs are the intersection of the proven receivers' numeric field sets, because the arm consumes one set and does not re-check which receiver a property belongs to. The PR's own commit message states a union would be "a wrong answer, not a weaker one", and nothing checks it.
This is harmless as shipped: with the knob off the fixpoint computes exactly what it computed before, so it cannot regress anything, and the gap suite confirms the off-path is a no-op. It is filed because the repo's own binding policy is explicit that an unexercised mode is a configuration nobody has verified — PERRY_GC_FORCE_EVACUATE was inert for every gc()-driven test (#6942/#6946), and the matrix's --pressure knob disabled the very path it was measuring (#7024).
What a test needs
Two proven receivers where a field is numeric on one only (must not be admitted) and a field numeric on both (must be admitted), with the knob on.
The obstacle is that nbc_order_enabled() reads std::env at the leaf, and perry-codegen's tests run in parallel, so an env-setting test races every sibling. Either isolate it (per_test_global!-style, as #10944 is doing for the asserted-counter class) or plumb the flag in as a parameter so the fixpoint can be exercised directly without touching the process environment — the latter is probably cleaner and removes the env read from the hot path as a side effect.
Refs #10777.
PERRY_L14_NBC_ORDER(landed in merge train 254,2dda7c1139/ #10929) is default-OFF, andshape_numeric_inputs()early-returns empty unless it is set:grep -rn PERRY_L14_NBC_ORDER crates/finds exactly four sites — the gate function, its doc comment, and the two cache registries (build_cache.rs,object_cache.rs). No test sets the variable.So the reorder the PR exists to perform is executed by no suite, and neither is the rule most likely to be wrong — that the shape inputs are the intersection of the proven receivers' numeric field sets, because the arm consumes one set and does not re-check which receiver a property belongs to. The PR's own commit message states a union would be "a wrong answer, not a weaker one", and nothing checks it.
This is harmless as shipped: with the knob off the fixpoint computes exactly what it computed before, so it cannot regress anything, and the gap suite confirms the off-path is a no-op. It is filed because the repo's own binding policy is explicit that an unexercised mode is a configuration nobody has verified —
PERRY_GC_FORCE_EVACUATEwas inert for everygc()-driven test (#6942/#6946), and the matrix's--pressureknob disabled the very path it was measuring (#7024).What a test needs
Two proven receivers where a field is numeric on one only (must not be admitted) and a field numeric on both (must be admitted), with the knob on.
The obstacle is that
nbc_order_enabled()readsstd::envat the leaf, andperry-codegen's tests run in parallel, so an env-setting test races every sibling. Either isolate it (per_test_global!-style, as #10944 is doing for the asserted-counter class) or plumb the flag in as a parameter so the fixpoint can be exercised directly without touching the process environment — the latter is probably cleaner and removes the env read from the hot path as a side effect.Refs #10777.