[bpamiri]

Summary

Part C of the approved performance slice. properties() read $get("resetPropertiesStructKeyCase") once per property inside its loop; it now reads the setting once at the top of the call.

 		local.rv = {};
 		local.propNames = propertyNames();
+		local.resetPropertyKeyCase = $get("resetPropertiesStructKeyCase");
 		for (local.key in this) {
 			...
-			if ($get("resetPropertiesStructKeyCase")) {
+			if (local.resetPropertyKeyCase) {

The flag is a call-local (local.resetPropertyKeyCase), exactly as the brief §C.2 prescribes — not an instance slot. That matters for correctness: model instances can be shared or long-lived, $get() honours per-tenant overrides, and a variables-scope flag would let two concurrent properties() calls on the same instance read each other's value mid-loop, and let a callback that re-enters properties() overwrite the outer loop's flag. A local has neither problem.

Semantics (unchanged)

• One read per properties() call; a set() between two calls decides the casing of the next call's keys.
• No cache and no invalidation surface — a deliberately rejected request-scoped $get memo (brief §C.3) is not used.
• No public function/argument changes, no engine branch, no changelog fragment (brief says none).

Regression test

Added to propertiesSpec.cfc: "reads resetPropertiesStructKeyCase once per call, so a set between calls is visible". It toggles set(resetPropertiesStructKeyCase = false) then true between two properties() calls and asserts the returned key spelling follows — the raw this-scope spelling when off, the canonical propertyNames() spelling when on. Both spellings are read from the engine at runtime so the test is valid on Adobe (upper-case this scope) and Lucee (lower-case), and it guards against any future per-request memo.

Measured impact

The microbench is small and inconclusive on this host; I am reporting it as such rather than claiming a win:

• $get("environment") costs ~1.6 µs/call on RustCFML (in-model and external), so removing 11–12 calls per properties() call is ~18 µs of work.
• A same-server A/B of old vs hoisted methods showed differences on the order of a few percent on RustCFML and within noise on Lucee (~80 µs baseline), but host variance exceeded the effect and the runs were not a rigorous multi-run, both-engine sample — so no number is claimed as a win.
• An early draft kept the flag in the variables scope; it was replaced with the brief's call-local for the thread/tenant and recursion reasons above. Correctness beats a sub-µs read, and the PR's own CI is the authoritative check.

Verification run on this change (head 7a25a6bc6)

• bash tools/test-local.sh model → 1502 passed (includes the regression test, green)
• bash tools/test-local.sh database → 189 passed
• tools/test-matrix.sh lucee7 sqlite passed during development (full suite); PR CI re-runs the engine matrix on this head.

Part C of the approved performance slice; PR A (#4631) is merged, PR B was dropped after its premise measured as a regression.