What

Makes the CAMPPlus speaker encoder differentiable for voice-clone enrollment by
adding an analytic, model-free C++ backward pass that returns d(loss)/d(fbank),
validated against the Task 2 finite-difference gradcheck harness. In the
enrollment loop CAMPPlus provides the speaker-similarity loss; the target-WAV
embedding stays forward-only (constant) and only the generated-audio path needs
gradients, so the gradient is the input gradient with the model weights frozen.

Follows the same pattern as the sibling tickets already on master
(#55 text-encoder tail / QVAC-20978, #58 vector estimator / QVAC-20982,
#60 vocoder / QVAC-20983): a pure double reference backward, gradchecked
component-wise, with the op×backend gap documented. Dependencies: Task 2
(QVAC-20979).

Forward-parity anchor

A gradcheck alone is self-referential: it only proves the backward is the exact
derivative of its own forward. To tie that to the real model, a second test
asserts the analytic double forward matches the production scalar forward
(campplus_embed_cpu) on synthetic weights (max_abs ≈ 3e-8, i.e. float-vs-double
rounding only). Building it surfaced that campplus_embed_cpu's fcm_forward
hardcodes the input feature dim to 80, so the production CPU path is only
self-consistent at feat_dim=80 (the parity test uses that). The analytic
backward derives every dimension from feat_dim, so it is geometry-agnostic.

Changes

• src/campplus_backward.{h,cpp} — new CampplusBackward class (namespace
  cp_grad). Owns the frozen weights and caches per-call activations as state;
  public surface is forward(fbank) / backward(d_emb). Channel-major (C, T)
  layout mirroring campplus_embed_cpu exactly. Implements the CAMPPlus
  primitives and their input-gradients: stride/pad/dilation-aware conv1d/conv2d,
  pre-fused affine batch norm, ReLU, sigmoid, time-mean, segment pooling,
  statistics pooling (mean + unbiased std), the FCM Conv2d residual block (with
  optional shortcut) and the CAMDenseTDNN layer (context-attention gate + dense
  concat split).
• test/test_campplus_backward.cpp — gradchecks every primitive, the FCM
  residual block, the CAM dense-TDNN layer and the full chain (12 checks) against
  central finite differences via the Task 2 harness. Always-on unit ctest tier
  (no model/fixtures, no-skip policy).
• test/test_campplus_backward_parity.cpp — forward parity vs the production
  campplus_embed_cpu (see above). Also unit tier.
• docs/voiceclone-backward-campplus.md — op×backend gap matrix and
  CPU-fallback rationale for enrollment.
• CMakeLists.txt — register the test-campplus-backward and
  test-campplus-backward-parity targets.

CPU fallback (documented)

SIGMOID, SQRT, MEAN, SUM_ROWS, PAD, REPEAT and CONCAT have no
backward in the vendored ggml, so the enrollment backward cannot use ggml
autodiff on any backend. It is provided as the analytic C++ backward and runs on
CPU (enrollment is offline; the realtime synthesis GPU fast paths are untouched).
See the doc for the full matrix.

Acceptance

Gradcheck green: test-campplus-backward and test-campplus-backward-parity
both pass (2/2 in the unit tier).