fix(bench): regenerate the official artifacts from real circuit-level measurements

The v0.7.0 artifacts reproduced the exact defect that got six pre-v0.7.0 ones
withdrawn (todo6 A1-03), and this time stamped _provenance's canned
METHODOLOGY_NOTE - which asserts estimate_ler_circuit_level scoring and
assert_comparable validation - into a file produced by neither:

* "LER" was `(H @ correction) % 2 == syndrome`, a syndrome-consistency check.
  It asks whether the decoder returned *a* valid correction, never whether the
  logical observable flipped, and read 0.000% for every decoder at every
  distance - the shape of a tautology, not a measurement.
* QECTOR was timed through a native batch_decode over the full shot count while
  PyMatching went through a Python `[dec.decode(s) for s in ...]` loop over as
  few as 20 shots, linearly extrapolated to 100,000 and written into
  `elapsed_sec` as if measured. The resulting "55x-108x faster than PyMatching"
  was mostly Python per-call overhead.
* chart_official_ler.png contained no measurements at all: hardcoded analytic
  curves `0.14 * (0.03/0.10)**((d+1)/2)`, constants picked per decoder by hand.

The generator now routes every decoder - QECTOR's backends, PyMatching and ldpc
alike - through ler.estimate_ler_circuit_level: one circuit, one decomposed DEM,
one detector/observable sample set per cell, one decode_batch resolver, scored
against the circuit's own logical observables. ler.assert_comparable gates the
rows before writing. Nothing is extrapolated; cells over the per-cell decode
budget are recorded as skipped with their measured probe rate and projected
cost, and are absent from the charts.

Measured, 77 cells (p=0.005, seed=1), largest shot count per cell:

    d=3   PyMatching 2,437,651/s LER 0.01891 | QECTOR Blossom 245,629/s 0.01891
    d=5   PyMatching   249,065/s LER 0.01596 | QECTOR Blossom   9,121/s 0.01596
    d=11  PyMatching    13,816/s LER 0.00647 | QECTOR Blossom      84/s 0.00700

QECTOR's Sparse Blossom reproduces PyMatching's decoding decisions exactly at
d=3 and d=5 (1891 and 1596 errors in 100,000, identical) - a real and checkable
accuracy result. On speed PyMatching leads at every distance measured. Both
statements are now in the artifacts.

Author: qectorlab