Research digest — Vendored libsvm + IQA coverage uplift (ADR-0952)¶

• Date: 2026-05-31
• Branch: test/vendored-libsvm-iqa-coverage
• ADR: ADR-0952
• Companion PR: #381 (libsvm parser audit, ADR-0889) — in flight

1. Trigger¶

Operator task brief (2026-05-31, verbatim translated to English):

Push test coverage on vendored libsvm (svm.cpp) + iqa paths that the fork actively uses. PR #381 (libsvm audit added 9 tests via test_svm_parser.c) already covers parser-rejection branches; do not overlap. Add 5–10 focused tests. Never modify upstream Netflix vendored content semantically — only add coverage. Document why if a path can't be tested without modifying vendored code.

2. Method¶

1. Build a clean coverage-instrumented tree from the worktree: meson setup build-cov core -Db_coverage=true -Denable_cuda=false -Denable_sycl=false
2. Run the fast suite to baseline coverage: meson test -C build-cov --suite=fast (49 tests, all pass).
3. Collect per-file coverage with gcovr: gcovr -r .. --print-summary --gcov-ignore-parse-errors=negative_hits.warn_once_per_file -f '.*svm\.cpp' -f '.*feature/iqa/'
4. Inspect uncovered line ranges to identify which functions were uncovered (not just lines); narrow the test design to those.
5. Add tests; rebuild; clear *.gcda; re-run fast suite; re-measure.
6. Diff coverage and document the per-file delta.

3. Baseline (master @ e81b13352, before this PR)¶

4. Gap analysis¶

4a. libsvm runtime (svm.cpp)¶

PR #381 adds test_svm_parser.c which targets parser rejection only: the eight if (!stream) / if (n > MAX_*) early-return paths plus unknown svm_type. That maxes out svm_load_model / svm_parse_model_from_buffer failure coverage but does not touch:

• svm_train (the SMO solver + decision-function construction)
• svm_predict, svm_predict_values, svm_predict_probability (Platt scaling + sigmoid_predict + multiclass_probability)
• svm_check_parameter (~12 rejection branches, plus the NU-SVC feasibility scan)
• svm_save_model (write side of the parser's read-back contract)
• The inspector family: svm_get_svm_type, svm_get_nr_class, svm_get_labels, svm_get_sv_indices, svm_get_nr_sv, svm_get_svr_probability
• svm_free_and_destroy_model (nulls the pointer; observable invariant)

4b. IQA helpers¶

test_iqa_convolve.c runs SIMD-vs-scalar bit-exactness on iqa_convolve only — and skips entirely when the host has neither AVX2 nor AVX-512 nor NEON. So iqa_convolve could be at 41% on a host with both AVX2 and AVX-512 and at 0% on an ARM CI runner that isn't NEON-tagged. No other IQA helper has any direct test.

iqa_ssim is reached only by float-SSIM extractor smoke tests, which all live under the enable_float=true gate and don't exercise the function with controlled inputs whose l/c/s outputs can be asserted.

5. Approach¶

Two new fast-suite test executables — observation-only against the public API. No vendored source is modified:

5a. core/test/test_svm_api.c (8 tests)¶

• test_check_param_accepts_default — happy path on the linearly separable 2-class fixture.
• test_check_param_rejects_unknown_svm_type — bumps svm_type to 99, asserts the "svm type" error string.
• test_check_param_rejects_unknown_kernel — same for kernel_type.
• test_check_param_rejects_bad_numeric — 10 separate sub-assertions for cache_size<=0, eps<=0, C<=0, shrinking != 0/1, probability != 0/1, RBF + gamma<0, POLY + degree<0, ONE_CLASS + probability=1, NU_SVC + nu=0, EPSILON_SVR + p<0.
• test_train_csvc_inspectors_and_predict — trains a C-SVC, inspects svm_type, nr_class, nr_sv, labels, sv_indices, check_probability_model (negative when probability=0), svm_get_svr_probability (returns 0 for classifier), predicts on side and − side, asserts svm_predict == svm_predict_values, drives svm_free_and_destroy_model and asserts pointer nulled.
• test_predict_probability_csvc — re-trains with probability=1, asserts check_probability_model == 1, calls svm_predict_probability, asserts probs sum to 1 and lie in [0,1].
• test_train_epsilon_svr — drives the SVR solver branch + the probA-only check_probability_model path + svm_get_svr_probability for the regressor.
• test_save_load_roundtrip — svm_save_model then svm_load_model; asserts inspector outputs match and predictions on the same query agree byte-for-byte.

5b. core/test/test_iqa_helpers.c (21 tests)¶

• 5 math_utils tests: _round (positive half + negative trunc-toward- zero asymmetry), _min / _max, _cmp_float (including rounding-by-scale boundary), _matrix_cmp.
• 5 KBND_* tests: in-bounds passthrough, negative reflect, positive reflect, replicate clamp, constant fallthrough.
• 6 filter tests: iqa_filter_pixel (NULL kernel + interior 3x3 box edge with REPLICATE border), iqa_img_filter (writes result, rejects NULL bnd_opt with rc=1, in-place when result==NULL).
• 3 decimate tests: factor-2 with no kernel + in-place + odd dimension (sw = w/factor + (w&1) — the half-pixel ceiling).
• 2 iqa_ssim tests: identical frames (ssim ≈ 1.0, all of l/c/s ≈ 1.0) random frames (asserts finite values + ssim < 0.99). These drive ssim_tools.c's scalar precompute / variance / accumulate fallbacks (since no SIMD dispatch is installed in the test process).

6. Hypothesis check — _round() asymmetry¶

Initial test draft assumed _round(-0.5) == -1 (round-away-from-zero on .5). Built and ran; assertion failed. Constructed a standalone verifier of the vendored body:

int _round(float a) {
    int sign_a = a > 0.0f ? 1 : -1;
    return a - (int)a >= 0.5 ? (int)a + sign_a : (int)a;
}

For a = -0.5: (int)a = 0, a - 0 = -0.5, -0.5 >= 0.5 is false, so returns (int)a = 0. The function is asymmetric — it rounds away from zero only on the positive side. Test was rewritten to lock the observed behaviour in as a regression assertion; the rounding rule becomes documented via the assertion.

Same hypothesise → check → act flow caught a wrong _cmp_float(1.234, 1.235, 2) expectation: 123.4 truncates to 123 but 123.5 adds the sign and becomes 124, so the two differ at digits=2, not agree. Test was rewritten with a verified pair (1.2 vs 1.21 @ digits=1) for the equality case.

7. Result¶

Re-measured after the two test files land + 29 new assertions pass:

Branch coverage on the same surface went from 8% to 54.6% — over half of the branch arms in the vendored bodies are now exercised.

8. Untested-without-modifying-vendored¶

Two surfaces remain uncovered and would require modifying vendored code to test:

1. svm_cross_validation — sits at ~30 lines, allocates a training subset and re-trains. Coverable without modifying vendored code, but the time cost (~5 retraining loops on a small fixture) pushes the test toward the slow suite. Defer to a follow-up.
2. OOM allocator hooks inside svm_train — Malloc / realloc failure paths are unreachable without an LD_PRELOAD shim or a vendored-side hook. Out of scope for an observation-only audit.
3. ssim_accumulate_lane.h SIMD-accumulator inline at 0% — only reached when iqa_ssim_set_dispatch installs a SIMD fn, which is gated by feature-build choices the test process doesn't pick up. Covered transitively when run inside the full extractor; not coverable from a standalone test without faking the dispatch.

9. Reproducer¶

git checkout test/vendored-libsvm-iqa-coverage
meson setup build-cov core -Db_coverage=true -Denable_cuda=false -Denable_sycl=false
ninja -C build-cov
meson test -C build-cov --suite=fast               # 51/51 pass
cd build-cov
gcovr -r .. --print-summary \
    --gcov-ignore-parse-errors=negative_hits.warn_once_per_file \
    -f '.*svm\.cpp' -f '.*feature/iqa/'

10. References¶

• req — operator task brief, 2026-05-31 (full text §1 above).
• core/src/AGENTS.md §10 — vendored libsvm fork-patch invariants that this test pass leaves untouched.
• ADR-0889 (in flight via PR #381) — libsvm parser audit; this work complements it on the runtime side.