Add --report flag to winml build for agent-friendly JSON output

Summary

Add a --report <path> flag to winml build that writes a structured JSON file capturing the full build result — stage-by-stage status, timings, artifact paths, and analyze findings — so downstream agents and automation pipelines can consume build outcomes without parsing Rich console output.

Motivation

winml build currently has no machine-readable output. All result data (stage timings, artifact paths, analyze node counts, autoconf patterns) is rendered through the Rich console and discarded. This blocks:

• Agent pipelines (e.g., winml flow) from reading build results to decide next steps
• CI/CD from asserting on specific stages without scraping terminal output
• Tooling integrations (foundryTK, VS Code extensions) from surfacing build summaries to users

The only partial exception is module mode, which writes a minimal module_summary.json — but it omits per-stage details and is not available in single-model mode.

Proposed Solution

Flag

winml build -c config.json -m ProsusAI/finbert -o output/ --report output/build_report.json

If --report is omitted, behavior is unchanged (no file written).

Report schema

{
  "schema_version": "1",
  "timestamp": "2026-05-06T10:23:00Z",
  "model_id": "ProsusAI/finbert",
  "config_file": "config.json",
  "output_dir": "output/",
  "success": true,
  "total_elapsed_s": 94.2,
  "stages": [
    {
      "name": "export",
      "status": "success",
      "elapsed_s": 18.4,
      "artifact": { "path": "output/export.onnx", "size_bytes": 438291456 }
    },
    {
      "name": "optimize",
      "status": "success",
      "elapsed_s": 31.7,
      "artifact": { "path": "output/optimized.onnx", "size_bytes": 412680192 },
      "analyze": {
        "iterations": 2,
        "converged": true,
        "autoconf_patterns": ["disable_layer_norm_fusion"],
        "node_counts": { "supported": 312, "partial": 4, "unsupported": 0 }
      }
    },
    {
      "name": "quantize",
      "status": "skipped",
      "reason": "QDQ nodes already present"
    },
    {
      "name": "compile",
      "status": "success",
      "elapsed_s": 44.1,
      "artifact": { "path": "output/compiled.onnx", "size_bytes": 109051904 }
    }
  ],
  "final_artifact": { "path": "output/model.onnx", "size_bytes": 109051904 },
  "reused": false
}

Failure case adds a top-level "error" field with the message, and the failed stage has "status": "failed".

Implementation notes

• stage_timings in _run_single_build already collects (name, elapsed | None) — extend it to carry status, artifact path/size, and analyze details
• Write the report in the finally block of _run_single_build so partial results are captured even on failure
• Module mode should extend the existing module_summary.json with the same per-stage detail rather than a separate file

Acceptance Criteria

• --report <path> flag added to winml build
• JSON written on both success and failure (partial results on failure)
• Report includes: schema_version, timestamp, model_id, config_file, output_dir, success, total_elapsed_s, stages, final_artifact, reused
• Each stage entry includes: name, status (success/skipped/failed), elapsed_s, artifact (path + size), and for optimize: analyze sub-object
• Module mode extends module_summary.json with per-stage detail
• Unit tests cover report schema for success, skip, failure, and reuse cases
• --report omitted → no behavior change (no file written, no performance impact)

Alternatives Considered

• Parse console output: Fragile, breaks on any Rich formatting change, not suitable for agents.
• Always write a report to output dir: Adds implicit side-effects; opt-in flag is cleaner.

Additional Context

The --report output is the intended data source for winml flow (see #442) so each pipeline step can surface a structured summary before prompting the user to continue or skip. It is also a natural data feed for the foundryTK Template UI to display build progress and results without re-running the build.