[D-sudoasd]

Raw DIC image sequences can now be analyzed through a headless Agent workflow with bounded parameter pilots, resumable per-frame computation, quality states and traceable scientific exports. The existing GUI, 1D extensometer, campaign workflow and ezDIC command/executable compatibility remain available.

The new straintrace agent CLI and straintrace_agent Python API cover project creation, input inspection, planning, analysis, cancellation/recovery, batch execution, verification, field plotting, ROI/path extraction and export. Native frame identity and missing-frame time gaps are preserved. Mechanics synchronization requires explicit clock/units/gap constraints; physical displacement and stress require supplied calibration and area. Unsupported values remain unavailable.

Field analysis adds consistent strain validity masks, Green–Lagrange and infinitesimal components/principal values, fixed-reference virtual gauges, sparse common-scale maps, position–time plots and labeled mechanics/gauge histories. CSV, safe NPZ checkpoints, strict JSON, content hashes and reports distinguish numerical quality from experimental validation. No third-party DIC source code was copied.

The reproducible examples/agent/create_demo.py --run workflow exercises known deformation, clock offset/unit conversion, bounded load interpolation, missing/corrupt frames, field extraction, checkpoint reuse, batch execution, source integrity and lock recovery. CI runs it through both an installed wheel and the Windows portable CLI.

Checkpoint reuse binds exact native/source/reference/time identity and the planned solver parameters; direct execution validates the plan checksum. Cancelled/incomplete runs cannot report a numerical pass, stale post-processing bindings are rejected before writes, runtime failures can retry, and repeated analyze/batch reuse a current plan and run. Windows transient replacement and strict cp1252 JSON output are covered.

Validation on runtime 9662c9ea and final example 8ad47cb:

• Independent review: approve, no blockers; the final Unicode console regression round-trips Chinese messages and paths through strict cp1252.
• Known-deformation end-to-end demo: GL Exx absolute error 0.00036095 and path-u error 0.018796 px; missing/corrupt frames, load interpolation, extraction, resume and integrity checks pass.
• Public Ncorr sample12: 12/12 frames, mask holes and ROI/path extraction preserved; source hashes unchanged. Operational validation only, no independent experimental truth.
• Two-sample 256-frame batch: all frames pass numerical gates; identical repeat retains plans/run IDs and all checkpoint hashes, with 256 reused frames and zero new DIC calculations. All 256 source hashes are unchanged.
• Source, installed-wheel and Windows portable workflows pass under cp1252; both packaged workflows also pass with Chinese input/output paths.
• Final full suite: 438 passed / 3 platform-dependent skips locally; 439 passed / 2 skips in GitHub Windows CI. All seven current PR checks pass, including installed/frozen Agent workflows, GUI smoke, locked benchmarks, Python distribution and JOSS artifacts/PDF.

Reproduce the complete synthetic experiment after installing the project dependencies:

python examples/agent/create_demo.py --output outputs/agent-demo --run --cli source

The workflow uses fixed-reference in-plane 2D DIC. Automatic ROI/gauge candidates do not confirm physical specimen geometry. Experimental uncertainty calibration, stereo/3D DIC and crack topology tracking are outside this change. A numerical quality pass does not approve a scientific interpretation. Long path scatter plots retain their selected samples; the field guide documents direct streaming CSV extraction for millions of samples without invoking that renderer.