codecaliper v0.2.2

• package 0.2.2 · spec 1.2.1 · grammars: tree-sitter-python 0.25.0,
  tree-sitter-java 0.23.5, tree-sitter-go 0.25.0 (binding tree-sitter 0.26.0)
• The instrument itself is unchanged; everything below is validation-lane
  work prompted by an external blind-review pass over the ICPC draft.
• validation/breadth/ now also measures the 119 Python snippets of the Dorn
  corpus under the Python grammar (new Dorn-Python row: 106/119 clean parse;
  the three pre-existing rows reproduce byte-identically). The Dorn-Java row
  is renamed from Dorn.
• validation/bw_faithfulness/ gains four probe scripts, all runnable from
  the tracked pins alone: annotator_sensitivity.py (leave-one-row-out over
  the 121 annotator rows: zero label flips, the 59/41 split and the 21/24
  sign-agreement count invariant), brace_repair_probe.py (adaptive brace
  balancing repairs 84/100, 94/100 with the scaffold, against 27 as written
  and 29 scaffold-only), uncertainty_probes.py (per-feature bootstrap CIs:
  18 of 24 exclude zero; paired tab-8-vs-1 contrast -0.23, CI [-0.52, +0.05];
  out-of-fold label stability across the 32 arbitration cells: max 16 flips,
  all in error-opaque cells) and nested_arbitration_probe.py (the decision
  rule re-run inside each outer fold: selection-aware accuracy 0.800
  [0.750, 0.860] against the non-nested 0.820 [0.770, 0.870]).
• New validation/crossparser/ lane: the parse anatomy re-measured with two
  architecturally different front ends (javac's JavacTask.parse and
  javaparser-core 3.26.2, sha256-pinned). Strict compilation-unit entries
  accept 0/100 snippets as written; entry-tolerance-normalized counts agree
  with the instrument (scaffolded 27/29 vs 29; balanced+scaffold 84/93 vs 94).
• New validation/audit/ lane: the two first-party readability tools
  recovered and audited (artifacts sha256-pinned, fetched at run time, never
  committed). The original Buse-Weimer tool's source counts a tab as FOUR
  columns in the indentation feature only (in-source comment "might add more
  than just 1?"), its snippet path is purely lexical, and its operator
  features are per-line character counts with no comment exclusion;
  re-counting under that semantics on the 100 pinned snippets yields Spearman
  +0.10 (comments included) against -0.24 (comments stripped), diagnosing the
  persistent avg_arithmetic_ops sign divergence as a counting-semantics
  artifact. Scalabrino et al.'s rsm.jar values every tab at zero columns in
  its BW features and returns NaN for truncated fragments in scoring mode.