qector-dem-pipelinelisted

# QECTOR DEM Pipeline The detector error model is the standard machine-readable description of a decoding problem (manual 14). The QECTOR pipeline accepts a Stim DEM (object or text), collapses it to a graph, applies per-mechanism priors and weights, and returns a faithful decoder. ## The two surfaces 1. **Library bench server (Provisional)** — `qector-research.dem_inspect` and `qector-research.dem_collapse_parallel`. A minimal Stim-style parser that is enough for the small fixtures the reference manual uses; it does not require Stim or the optional direct-wheel `dem` module. 2. **Optional direct-wheel `dem` module** (Provisional, manual 16.4) — `dem.from_stim(text)`, `model.collapse_to_graph()`, `model.make_decoder('blossom')`. Verify the exact API on the target device by introspecting the installed wheel. ## The collapse rule (manual 14.1) Parallel mechanisms between the same detector pair are merged into one edge. For two mechanisms `p1` and `p2` the combined probability is p = p1 (1 - p2) + p2 (1 - p1) (the independent-XOR rule, also what `stim`'s `detector_error_model(decompose_errors=True)` produces). The merged edge keeps the **observable set of the more likely member**. The matching decoders only ever use the lowest-weight edge between two detectors, so the collapse is exactly what PyMatching does, and it preserves logical accuracy on the tested workloads. For `n` mechanisms the cumulative form is p_combined = (1 - prod_i (1 - 2 p