Index of the seven small pull requests from the FHE-MoE inference study (#34-#40): what each one fixes, and a suggested order

[seyfal]

We opened seven independent PRs while building a 32-layer CKKS mixture-of-experts inference chain on FIDESlib 2.1.3 (H200, OpenFHE 1.2.x front end). Each is small (5-46 changed lines), touches one concern, has a reproducer in its description, and merges cleanly against main in any order (#34 and #40 both edit src/CudaUtils.cu but in different functions). This issue is only a map, so a reviewer can pick an order and see the whole batch at once.

PR	area	one line	lines
#35	Accumulate/Broadcast	reject bStep values that hang or give wrong sums (validation + tests)	+46/-2
#36	Ciphertext ops	throw when an operand cannot be adjusted to the ciphertext instead of silently skipping the operation	+28/-7
#37	Bootstrap	reject inputs with fewer limbs than their NoiseLevel needs instead of returning garbage	+26/-3
#38	rotate_hoisted	copy slots to the result in the non-fused path; document ext	+9/-4
#39	LimbPartition::multPt	use the top limb of the current level (FIXEDMANUAL multPt with rescale was wrong below the top level)	+5/-2
#34	CUDA utils	create the CUDA stream pool lazily instead of from a static initializer	+11/-1
#40	CUDA utils / Context	clearAuxilarPoly: one device synchronization instead of an event pair per pooled buffer (12.0 ms -> 0.6 ms per call on H200, 2.1 GB pool)	+15/-1

Suggested order: the wrong-result fixes first (#39, #38), then the fail-loudly guards (#35, #36, #37), then the two CUDA-utils changes (#34, #40). None depends on another; if you would rather review them as one or two combined PRs, say so and we will restructure them.

How they were found: every item comes from a mismatch between the GPU chain and a host reference (routing-aligned numpy chain) or from a hang, and each description names the input that triggers it. Happy to add tests or split further.

Fable 5.1 on behalf of Seyfal