//tests/Examples/lattigo/ckks/bicyclic_matmul:bicyclic_matmul_test is flaky

[j2kun]

Disabled until we figure out why.

[github-actions]

This issue has 2 outstanding TODOs:

• tests/Examples/lattigo/ckks/batch_matmul/BUILD:16: enable it once the flakiness of bicyclic layout is resolved
• tests/Examples/lattigo/ckks/bicyclic_matmul/BUILD:16: disable until flakiness is resolved

This comment was autogenerated by todo-backlinks

[wtd2]

AI suggested the following to fix this issue:

Root cause

The failure is deterministic, not runtime flakiness: at current main
(9ba0866) the generated test binary fails 50/50 local runs, all slots
wrong with large per-run-random values.

lattigo-alloc-to-inplace reuses a dead degree-2 ciphertext as the
output buffer of an in-place Add whose operands are both degree-1.
In the generated bicyclicmatmul_lib.go:

err54 := evaluator.Add(ct6, ct9, ct3)    // ct3 := deg2 + deg2   (valid deg-2 partial)
err55 := evaluator.Add(ct2, ct3, ct2)    // ct3's value consumed into ct2
err56 := evaluator.Relinearize(ct2, ct2)
err57 := evaluator.Add(ct12, ct16, ct3)  // deg1 + deg1 -> reuses ct3 as buffer  ← HERE

Lattigo sizes the output of a binary in-place op over the receiver's
current degree too (lattigo v6.1.0, core/rlwe/evaluator.go,
InitOutputBinaryOp):

degree = utils.Max(op0.Degree(), op1.Degree())
degree = utils.Max(degree, opOut.Degree())   // receiver participates

and evaluateInPlace only writes components 0..minDegree. So at
err57 the reused buffer keeps its stale c2:

ct3.Value = [x0, x1, x2]                  // dead leftover from err54
Add(ct12, ct16, ct3):
  output degree = max(1, 1, 2) = 2        // receiver keeps 3 components
  ct3.Value = [a0+b0, a1+b1, x2]          // x2 never touched
decrypt: (a0+b0) + (a1+b1)·s + x2·s²  =  correct sum + garbage

The stray x2·s² is essentially uniform mod Q (the old c0/c1 that
balanced it were overwritten), hence all slots garbage (±1e4 at Δ=2^45),
different each run because keys are fresh per run.

Confirmed by decrypting every intermediate in an instrumented copy of
the generated code: values are clean (max ≈ 0.17) through err56, jump
to ≈ 4.5e4 exactly at err57.

Proposed fix

Before emitting an in-place Add/Sub/Mul whose buffer is not one of
the operands, emit:

buf.Resize(max(lhs.Degree(), rhs.Degree()), min(lhs.Level(), rhs.Level()))
// Mul uses lhs.Degree()+rhs.Degree() (tensor-product degree)

The emitter already guards exactly this hazard for Rotate
(LattigoEmitter.cpp emits
inplace.Resize(input.Degree(), input.Level())) — Lattigo's Rotate
errors loudly on degree mismatch, so that case got fixed; Add/Sub/Mul
corrupt silently. This adds the analogous guard to the CKKS and BGV
in-place printers.

Validation at main 9ba0866: bicyclic_matmul_test 0/50 → 50/50
(plus a 20/20 rerun), bgv/binops e2e and the emitter lit test still
green.