test(ai): coverage push — 2 new test files lifting online_trainer + feature_extractor from 0%/partial to 85%/94%

[lusoris]

Summary

• Adds ai/sidecar/tests/test_online_trainer.py (30 tests) — lifts online_trainer.py from 0% → 85% coverage.
• Adds ai/tests/test_feature_extractor_unit.py (26 tests) — lifts feature_extractor.py from partial → 94% coverage.
• Adds changelog.d/added/ai-sidecar-online-trainer-coverage.md fragment.

Both modules were identified as the highest-LOC, lowest-covered targets in the ai/ subtree.

Type

• test — test-only

Checklist

• Commits follow Conventional Commits.
• make format && make lint is green locally (black, isort, ruff pass; semgrep-local hits pre-existing io_uring ENOMEM on dev machine — CI container is unaffected, same failure as PR fix(api): replace ISO-reserved __VMAF_*__ include guards with LIBVMAF_*_H (Round 27 A.1) #481 / docs(api): warning-clean doxygen build for libvmaf public C API #457 / docs(api): warning-clean doxygen build for libvmaf public C API — recovery of #457 #484).
• All 56 new tests pass: pytest ai/tests/test_feature_extractor_unit.py ai/sidecar/tests/test_online_trainer.py -v.
• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.

Bug-status hygiene

no state delta: test-only change, no bugs opened or closed.

Netflix golden-data gate

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.

Cross-backend numerical results

no cross-backend diff needed: test-only, no C source touched.

Deep-dive deliverables

• Research digest — no digest needed: trivial test-only coverage push.
• Decision matrix — no alternatives: only-one-way fix (write missing tests).
• AGENTS.md invariant note — no rebase-sensitive invariants: test files only.
• Reproducer / smoke-test command — see below.
• CHANGELOG fragment — changelog.d/added/ai-sidecar-online-trainer-coverage.md.
• Rebase note — no rebase impact: test files only, no C/public-header/upstream-mirror change.

Reproducer

cd ai
python -m pytest tests/test_feature_extractor_unit.py sidecar/tests/test_online_trainer.py -v
# Expected: 56 passed

Coverage delta

Module	Before	After
ai/sidecar/online_trainer.py	0%	85%
ai/data/feature_extractor.py	~40% (registry/resolver only)	94%

What is tested

test_online_trainer.py (30 tests)

• _build_fallback_model — output shape, accepts different n_features
• _load_base_model — empty path fallback, absent file fallback, PyTorch state-dict load, corrupt file fallback
• _write_sha256_sidecar — file created, valid 64-char hex, known SHA-256 value, atomic write (no .tmp leftover)
• OnlineTrainer.__init__ — construction, checkpoint dir creation, custom buffer capacity
• OnlineTrainer.ingest — accumulation below batch size, buffer fill, batch triggers gradient step, step count increments, checkpoint path returned, SHA-256 sidecar present
• OnlineTrainer.status — all keys present, checkpoint_counter increments
• _handle_connection — valid message returns ok, job_id echoed, missing features key returns error, malformed JSON returns error, empty line skipped, multiple messages on one connection
• run_server — starts, accepts connections, stops cleanly and unlinks socket

test_feature_extractor_unit.py (26 tests)

• _ensure_binary — raises RuntimeError when absent, passes when file exists
• _lookup — direct key, integer_ prefix fallback, absent returns None, direct preferred over integer_ prefix
• _run_vmaf_json — argv composition (flags, pix_fmt, bitdepth), feature extractor flags, temp file cleanup on subprocess error
• extract_features — happy path with 2 frames, NaN for missing metrics, integer_ fallback, empty frames, raises when binary absent, respects VMAF_BIN env
• FeatureExtractionResult — JSON round-trip, dtype preserved
• aggregate_clip_stats — shape is 4×n_features, correct mean, p10 < p90, NaN ignored, all-NaN propagates, empty frames returns NaN vector, raises on 1-D input, output is float32, single frame

🤖 Generated with Claude Code