v0.3 — Craft & Blast-Radius axes, reliability, and honest scoring

v0.3 — Craft & Blast-Radius axes, reliability, and honest scoring

v0.3 adds two new scoring axes (Craft, Blast radius), matures the
pairwise craft judge into a signal you can trust at small samples, makes
deterministic correctness real, and hardens the whole pipeline against the ways
it can quietly lie. 18 PRs since v0.2.

Design throughline: every scoring input is deterministic and diff-derivable
(re-derivable by hand from the saved diff) or a fail-closed judge call that
degrades to a tie/unknown — never a fabricated preference. The harness persists
each cell's diff, so slop/craft can be recomputed offline from a finished run
with no rerun.

New scoring axes

• Craft axis — deterministic "slop" metrics over the diff (duplication
  windows, churn ratio, TODO/debug/commented-out residue, test-tamper) plus an
  LLM judge scoring the qualitative residual, and a cross-bundle pairwise
  A/B judge for ranking. (#8, #27)
• Blast radius — a task or campaign link may declare expectedSurface
  globs; the harness mechanically computes out-of-scope excursions and the
  judge classifies each (necessary / defensible / overreach / adversarial). Any
  adversarial excursion disqualifies the cell from craft aggregation. Now armed
  on a real task. (#8, #7)

Craft — matured into a trustworthy signal

• Two new craft dimensions: documentation and testing — scored on
  value, not volume. Proactive docs and meaningful tests now count for an
  agent instead of being ignored or (in the case of docs) penalized as
  overreach. Reconciled with economy so they're never double-penalized as
  verbosity. (#38, + testing dimension)
• Severity-weighted pairwise verdicts — a soundness-implicating win
  (catching a real defect: an open redirect, a missing guard) counts as 3
  stylistic wins, so a caught vulnerability is no longer outweighed by a naming
  or import nit. Fail-closed to ordinary weight. (#35)
• Composite Craft Score — round(100·(0.7·winRate + 0.3·SlopHealth)) using
  a head-to-head macro-average win rate (beating the weakest variant no
  longer inflates a score) with a confidence layer: thin samples render
  ≈ / ⚠ low-confidence instead of false precision, and a shutout or large
  gap still separates. (#27, #34)
• Both-order pairwise de-biasing — opt-in --pairwise-both-orders judges
  each pair in both A/B seatings and combines (a verdict that flips on swap
  resolves to tie), cancelling judge position bias deterministically instead of
  relying on randomization at small N. (#36)
• Slop measures agent-authored production code only — doc files, test files,
  and generated lockfiles (package-lock.json, yarn.lock, …) are excluded
  from the code-hygiene metrics (they're rewarded or handled elsewhere, not
  slop). Duplication now carries per-file evidence for auditability;
  testTamper still guards test-weakening. This closes a false-positive where
  npm's lockfile duplication crashed the SlopHealth of any variant that added a
  dependency. (#43, #45)
• testTamper counts net assertion loss — a legitimate assertion rewrite
  forced by a schema change no longer false-positives as tampering. (#6)

Correctness — made real

• testCommand / judgeOnly — tasks with a runnable in-container harness
  now produce a real deterministic pass/fail verdict; tasks that genuinely
  can't run tests in-container are explicitly judgeOnly (by design, not
  silently empty) and marked distinctly in the Correctness table. (#22)
• Honest signposting — a loud warning when a matrix has scored cells but
  zero deterministic verdicts; the executor log now prints the real test
  verdict, not mere test-file presence. (#9)
• Hardened judge parse — recovers fenced / prose-wrapped / trailing-comma /
  truncated JSON (last-parseable object wins, so a leading example can't shadow
  the verdict) to cut spurious cell exclusions; fail-closed on true garbage. (#12)

Memory

• Held-by-abstraction (✓A) is now earnable — the campaign induces a
  reusable helper so a bundle that generalizes a convention grades ✓A (the
  gold-standard memory result) while a literal re-emitter grades ✓L. A
  reproduced false-✓A (a drift link graded gold via a shared domain noun) is
  closed by decoupling the linkage-exclusion vocabulary from appliesIf and a
  token-boundary declaration-only harvest. A ✓A headline callout now fires for
  campaign-earned wins. (#13, #28, #37)
• memory-registry flagged non-discriminating (✝) — a memoryless bundle
  recovers the rule from repo/task context, so a hold there is not a memory
  win. (#14)
• Drift vs trap — the campaign summary now splits cumulative adherence into
  held / drift / trap / unknown (different failure modes, different costs). (#15)

Reporting & observability

• Reliability — with --repeats N, per-cell min/mean/max + σ dispersion
  across the three major axes and per-dimension craft ranges. (#19)
• Sparklines — campaign adherence trajectory and per-task efficiency. (#21)
• Cross-task insight — a synthesized narrative callout derived from the
  behavioral/efficiency data already in the report. (#18)
• --focus <axis> — render a focused report for one concern
  (correctness / memory / craft / efficiency / blast-radius); the JSON
  payload stays complete. (#20)
• Token accounting — the Efficiency input-token column is labeled
  (uncached); it's the CLI's uncached delta, not total context cost. (#11)

Fixes rolled up

#6 #7 #9 #10 #11 #12 #13 #14 #15 #16 #17 #18 #19 #20
#21 #22 #28 #35 #37 #38 #43 #45 — see the v0.3 milestone.