docs(research): reuse-decline, specified precisely — and the two cross-context smells we do not measure

[joyful-ii-V-I]

An investigation note, not a feature and not a plan. No src/ change.

Three of --quality-delta's ten kinds — verbosity, duplication and reuse-decline — were shaped by Tsantalis, Zhu & Rigby, AI-generated code smells (arXiv:2605.02741), the design already recorded at docs/LINEAGE.md L84. Of the three, reuse-decline (new-clone-of-reused-helper) is the only one that is ours rather than a standard metric, so it is the only one worth putting in front of anyone to critique. Full write-up: docs/research/ai-smells-reuse-decline.md.

(Base note: measured against main at 755f9026, before 0.6.2; the diff still merges cleanly onto today's main, 15a20855, tag 0.6.2.)

Headline finding. This branch writes down precisely what reuse-decline computes, proposes two unmeasured smells from the same line of work, and measures all three on real trees — including two unflattering results. Cross-file share of clone groups is 57.1% in this repository and 66.7% pooled over a 90-repository corpus, but the per-repository median is only 37.3% across a 0%–94.5% spread, so the pooled figure is not a number to threshold on. Replaying the kind over 150 commits of this repository fires it exactly once, and that one row is a shell-gate harness clone that the sibling duplication kind exempts and reuse-decline does not — the same asymmetry the ack ledger corroborates: 43 of 62 accepted findings sit at fan-in 3 or 4, one notch above kReusedHelperMinFanin, a threshold that was reasoned rather than fitted.

A real src/ finding, written up and deliberately not fixed here (this lane's scope is docs/bench only): duplication's reporter skips all-test-script clone groups and demotes recognized idioms to minor; reuse-decline's reporter, reading the same clone vectors, does neither. It wants its own lane.

This is a first pass with a lot further to push. §4 of the document lists, in our own words, the six questions we would most like to be told we have wrong — the fan-in threshold first, then the two proposed-but-unmeasured smells (clone-group locality, and a retrospective "one rule living in N files" read). Corpus A (90 shallow-cloned checkouts) blocks any future history-based measurement on it, which is stated plainly rather than worked around.

Our working premise across this line of investigations: algorithmic, deterministic checks applied while an AI writes are the practical way to keep code sound at the speed AI now writes it — no label, no vendor claim, a check that either fires or it doesn't.

🤖 Generated with Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}