Native duplication defaults

Measured at a16d4372376bf18eea3dff4cbe85e15fc68e7fc0 with jscpd 5.4.2 on macOS arm64. The scans include native tests and Objective-C headers. They respect .gitignore, which excludes Swift build caches, Android build output, and packaged helpers. No source exclusions were added.

Reproduce the comparison

pnpm install --frozen-lockfile
pnpm native-dupes:prototype

Each variant scans apple/, android/, and linux/, then scans the two TypeScript tests named in the original report. JSON reports and summary.json go to .tmp/native-dupes-prototypes/. The prototype fixes the worker count at two. The table records this source revision, not a target for later revisions. Percentages are jscpd's duplicated-line metric, not unique lines to remove. Files below the token minimum do not contribute to its analyzed-file count.

Mode	Tokens / lines	Native clones	Duplicated lines	Duplication	Duo pair clones	Duo panel fragments
mild	50 / 5	85	873	1.80%	4	0
weak	50 / 5	82	857	1.77%	4	0
weak	70 / 5	18	228	0.47%	2	0
weak	100 / 8	0	0	0.00%	0	0
weak	30 / 3	401	2971	6.12%	16	0
weak	20 / 3	860	5435	11.20%	31	2
weak + consistent identifier normalization	50 / 5	196	2150	4.44%	5	0
weak + gaps up to 8 lines	30 / 3	384	2960	6.10%	13	0

The measured native runs took about 170–195 ms each, including Node startup, using two workers. This is a single timing sample per variant, not a performance benchmark. Linux execution belongs to the Repo Guards & Compatibility CI job.

Selected policy

Use weak mode with a minimum of 50 tokens and 5 lines. Weak mode removes whitespace and comments from matching. The default mild mode adds three findings without improving either motivating example. At 70 tokens, all seven Java findings disappear, including repeated snapshot capture and error response handling. At 100 tokens and 8 lines, the native report is empty. At 30 tokens and 3 lines, imports, small guards, and repeated test setup dominate a much larger review backlog. Identifier normalization also expands that backlog and still misses the Duo panels.

The selected scan analyzes 218 files and reports 82 clones: 68 Swift, 7 Java, 7 Objective-C, and 0 Python. There are no Kotlin source files in the measured trees; an executable fixture proves Kotlin detection. The Objective-C extension mapping includes .h, because those headers contain executable geometry code as well as declarations. Five header findings are declaration/implementation pairs. Existing findings also include deliberately similar tests and Swift package manifests. The baseline records these findings without declaring them defects or requiring extraction.

jscpd 5.4.2 has a built-in fingerprint baseline, introduced in 5.1.0. A custom comparison script or percentage ceiling is unnecessary. The baseline records content fingerprints and clone multiplicities. It tolerates line shifts and fails on a third copy of an accepted pair. A new clone fails even if another clone is removed and the total stays equal. The executable policy tests cover each language, headers, line shifts, extra copies, replacement clones, missing baselines, and empty scans.

The rollout starts with a report-only measurement and saves that result in native-dupes-baseline.json. check:native-dupes then rejects new fingerprints or increased multiplicities. It scans all native files rather than filtering by changed files. This keeps unchanged files available as comparison targets and needs no Git history in CI. check:affected selects it for native source changes; configuration and baseline changes select the full check set. --changed is unsuitable for PR enforcement because it filters by working-tree status, which is empty after a commit. The normal check never updates the baseline.

Motivating cases and limits

The closed-Duo fixtures differ at the measured revision. display-inventory.test.ts includes bounds, orientation, and unique IDs absent from simctl-facade.test.ts. At 50 / 5, jscpd finds shared imports and the separate single-panel fixture, but no shared closed-Duo panel fragment. At 20 / 3, it finds two 26-token panel fragments, at simctl-facade.test.ts:36–40 and 46–50 against display-inventory.test.ts:76–80 and 93–97. It also reports 860 native clones. Enabling gap merging cannot recover fragments below its seed token minimum. TypeScript remains owned by Fallow; this experiment does not add a second TypeScript enforcement policy.

No matrix variant finds a clone in the checked-in Fold.m. A separate 10-token / 2-line probe finds two tiny repeated fragments (lines 165–167 against 230–232, and 214–216 against 218–220). Those are not evidence that it would catch the historical seven-line case; that original duplicated revision was not replayed. Keep small or reimplemented logic in code review. Renamed or partially edited copies can evade the selected exact-token policy. Reported blocks need human assessment, especially test setup and Objective-C declarations.

Inspect or change the baseline

Run pnpm native-dupes:report to inspect findings without failing on new clones. Both report and check write .tmp/native-dupes/jscpd-report.json; each clone has an isNew field. CI uploads the check report as native-dupes-report even when the check fails.

When a check reports a new clone, compare both locations and remove the repeated block when they share an owner. If retaining duplication is intentional, review its source before running pnpm native-dupes:baseline. Review the baseline diff together with the source change. Do not regenerate it merely to clear a failing check. A detector upgrade or a change to matching options also needs a fresh comparison and baseline review. Missing baselines and empty scans fail closed.