fix(build): honor CLI optimization bypass

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

Summary

CLI-only control-correctness fix; no recipes, model-specific logic, new graph optimization pass, or speedup claim.

• Consume skip_optimize at the actual single-model HF and direct-ONNX CLI sinks, then pass CLI skip OR config.skip_optimize to the existing _run_optimize_stage helper.
• Retain the helper and its intermediate-copy/config behavior. The existing shared loop suppresses optimization and analyzer/autoconf re-optimization when bypass is requested.
• Preserve configured raw-model FP16/static quantization, compilation, and finalization. Direct-ONNX prequantization remains graph-driven (QDQ/QOperator); an ordinary bypass does not mark a raw model prequantized or clear its quantization config.

Exactly two changed paths: the CLI build command (+4/-1) and its regression tests (+252). No library, recipe, dependency, lockfile, or skill changes.

Validation

• Reviewed and shipped commit: 4da82cde6c93894199f00e108f80be13f5dfe3bd; parent/base: dcad2f4a29cb34af71d7b138bcafc4276a70a4ed. Fetched origin/main still matched this base before shipment; no rebase or new commit.
• Retained causal regression evidence: 14 failed / 26 passed -> 40 passed, with identical selected test identities. Retained committed affected-scope JUnit/logs: 358 passed, 0 failures/errors/skips; rechecked against the exact reviewed source/test blobs and case set.
• Fresh shipment check at the same commit: 40 passed, 115 deselected in 8.54 s, with no failures/errors/skips and the same 40 regression identities. Read-only Ruff check, Ruff format check, and Git whitespace checks passed.
• Explicit Python 3.11.16, verified worktree source origins, existing package/overlay environment, offline mode, and no new dependencies. The tests use actual Click dispatch and stage helpers with tiny pytest-generated raw/QDQ/QOperator graphs; expensive payloads and host/UI boundaries are mocked.
• Independent local assessment: LOCAL_READY, no critical/important introduced-code defect. Nonblocking artifact-finalization test-hardening feedback remains; this is not a GitHub approval. Independent post-PR review is still required.

Scope and limits

• Composite CLI-flag fan-out is covered. Array/module CLI mode, outer composite config-only inheritance, and optimized-bundle generic controls are pre-existing exclusions, not fixed or claimed supported here.
• No full-repository suite, real-model export/calibration/inference, numerical-quality, provider/hardware, or performance validation is claimed. This PR does not establish unrelated VitPose recipe quality or latency, and does not certify FP16/static-quantized model outputs.

Overlap and integration

Overlaps the CLI portion of #1370 by ssss141414 (observed head 3cd22269dfbfaac8baef63cb221c44c9e803af91). That draft combines multilingual recipes with a repair that bypasses the entire optimize-stage helper; this standalone fix instead preserves the existing helper contract.

Proposed disposition, if this fix is adopted: use this as the canonical CLI repair in place of #1370's overlapping code/test change, while keeping its model recipes separate or adapting them to depend on the canonical repair. This acknowledges the prior overlapping work; it does not claim agreement from that author or declare #1370 superseded. #1370 has not been edited, commented on, closed, rebased, or pushed by this shipment.

Coordinate ownership before landing either overlapping implementation; do not merge both unchanged. Keep this PR draft pending independent post-PR review and integration coordination.