[colbymchenry]

Summary

When a project keeps copies of the same code in several folders, like Android build flavors that each hold one package, or several example apps that each carry the same packages, a call now reaches the copy in the caller's own folder. In a deep tree it used to take whichever copy was indexed first, and the right pick was understated as a proximity guess.

Root cause

findBestMatch scores directory proximity at 15 points per shared leading folder, capped at 80, so two copies that both share six or more folders with the caller tie, and the first one indexed wins. In matchByExactName the winner's confidence came from proximity alone (0.7, or 0.4 in a shallow tree), though among copies of one symbol the name is never in doubt, only which copy.

Reproduced on main with a realistic layout: two Kotlin flavor source sets platform/apps/mobile/android/app/src/rc2/… and …/src/rcpro2/… holding one package. Both flavors' run() linked to rc2's compute(), and rcpro2's new Util() in Java to rc2's class.

What came from the contributor, and what we changed

From @danusha2345 (#1936): closestCopy and sharedDirDepth, the tie-break in findBestMatch (only among candidates that tie on the full score, and only when every one is a copy: same qualified name, kind and file name), the 0.8 confidence, the Python tests, and the CHANGELOG bullet.

Ours, on top:

1. Rebased onto fix(resolution): score a decisive same-file match as strong #1938 (merged as fix(resolution): a lone same-file definition scores as a strong match #2874), which edits the same confidence line. Resolved to keep both: the file's own decisive definition 0.9 first, then the caller's own copy 0.8, then proximity. The contributor's later force-push (5abdf25, onto current main) resolves that line the same way and is otherwise the patch we cherry-picked, so nothing more needed folding in.
2. Copies are judged among the candidates findBestMatch weighed. It skips other languages whenever a same-language candidate exists, so the tie-break already ignored them, but the 0.8 check didn't: a Python compute in the same repo kept two Kotlin copies at 0.7. In playframework this raised 13 more right picks (each sample project's own models.User and Application methods).
3. A match of the caller itself keeps the proximity score, as fix(resolution): score a decisive same-file match as strong #1938 does for a same-file match: recursion can't be told apart from a call that lost its receiver.
4. Realistic tests: the PR's Python fixture calls compute() without importing it, which Python can't do. Added Kotlin flavor copies (bare call in one package) and Java flavor copies (new Util()), plus the self-match case.
5. CHANGELOG: moved next to fix(resolution): score a decisive same-file match as strong #1938's bullet and reworded, with Thanks @danusha2345.

Validation

Tests. __tests__/same-module-copy-resolution.test.ts, 6 cases: 3 fail on main on the targets themselves (Python, Kotlin and Java copies), 2 fail on the PR as written, re-checked on its current head 5abdf25 (the Kotlin confidence and the self-match), all 6 pass here.

A/B, main (with #2874) vs this change, fresh indexes, site by site, 21 repos in 11 languages (flask, click, fastapi, gin, cobra, etcd, ripgrep, jq, Slim, serilog, eShopOnWeb, nowinandroid, okhttp, http4s, playframework, spring-modulith, express, zod, bulletproof-react, excalidraw, vben):

repo	edges	added	removed	retargeted	confidence only	via a framework rule before
spring-modulith	10,048	0	0	3	15	14 (+3 retargeted)
playframework	61,076	0	0	1	38	0
http4s	25,631	0	0	0	21	0
etcd	42,643	0	0	0	21	1
vben	20,184	0	0	0	10	0
bulletproof-react	4,420	0	0	0	3	0
okhttp	31,980	0	0	0	1	0
the other 14	130,768	0	0	0	0	0
total	326,750	0	0	4	109	15 (+3)

Triage:

• The 4 retargets are right. spring-modulith's spring-modulith-example-full tests now reach that example's own OrderManagement and OrderCompleted, where Spring's naming rule (0.7) had taken the epr-jdbc example's copy (3). playframework's shutdown-happy-path scripted test now instantiates its own src/main HomeController, not the changes/ copy its script swaps in later (1, the case the PR describes).
• The 109 confidence raises keep their target, each onto the copy in the caller's own module or app: every spring-modulith example's own classes, http4s's client tests reaching the client middleware and server tests the server's, each vben app's main.ts reaching its own bootstrap.ts, each bulletproof-react app's mocks reaching its own db.ts, etcd's pkg/proxy test reaching pkg/proxy. In Go, qualified names carry no package, so same-named functions in same-named files of different packages count as copies; the own-package pick is what a bare Go call means anyway. 4 of the 109 are wrong targets main already had at 0.7: vben's defineConfig calls from import { defineConfig } from 'tsdown' link to the folder's own src/index.ts namesake, because tsdown is declared only in the root package.json (a pnpm catalog: entry) and so isn't recognized as an outside package from a workspace package. That gap predates this change.
• The 0.8 against framework rules (0.7–0.75): in 18 sites it now beats one (spring-modulith 17, etcd 1); 3 are the retargets above, the rest pick the same target.
• After rebasing onto fix(play): a Class.method reference reaches a member of the class it names #2875 (Play's resolver), playframework was indexed again with the new main and this branch: the same 1 retarget and 38 raises, nothing else.
• Kernel parity doesn't apply: TypeScript resolution only, same kernel binary in both arms.

Cost, index process CPU time, interleaved, median of 3: okhttp 34.8 s vs 34.8 s, etcd 36.5 s vs 36.5 s.

Suite on this commit (CODEGRAPH_KERNEL_EXPECT=1 npx vitest run --maxWorkers=4 --minWorkers=1, Windows): 760 files passed, 6 skipped; 9,032 tests passed, 83 skipped, 0 failed.

Lands #1936 by @danusha2345.

🤖 Generated with Claude Code