[colbymchenry]

Summary

__tests__/explore-cross-call-dedup.test.ts › "a second call against a real index › never re-sends a line it already sent, and points at every line it withholds" has failed on every LF checkout (Linux, macOS, CI) since #2655. It kept passing on Windows only because core.autocrlf writes the fixture with CRLF. This makes the case's expectation match the range algebra's deliberate re-send of a short held run, and makes the suite index the same fixture bytes on every checkout. Test only: what agents receive doesn't change.

Root cause

• Since fix(go): a struct links the types its fields are declared with #2655 the payroll fixture's RunResult struct (internal/usecase/payroll/cycle.go lines 31-37) references Payslip through its Payslips []payroll.Payslip field, so call 1 renders it.
• Call 2 renders the file's head, and that 7-line held run comes back inside it. dedupeRange keeps a covered run shorter than EXPLORE_DEDUP.MIN_COVERED_LINES (8) in the emit set on purpose: rule 3 of the algebra, "cut chunks, not slivers", pinned by the unit test "withholds nothing when the overlap is smaller than a chunk worth pointing at". The handler case counted any repeated line as a duplicate.
• Why Windows passed: explore splits a file on \n, so on a CRLF checkout each line keeps its \r and the per-file character budget counts it. cycle.go then rendered fewer lines, and the case never met the short run. The suite tested a different scenario on each platform.

Fix

• The case groups the repeated lines into consecutive runs and fails on a run of EXPLORE_DEDUP.MIN_COVERED_LINES lines or more. The threshold is read from the module, so the two can't drift apart. A shorter run is the sanctioned re-send, beside the anti-abandonment restore the comment already named.
• beforeAll gives its copy of the fixture LF line endings (normalizeLineEndings, a byte-exact latin1 rewrite of \r\n only), so every checkout indexes and renders the same bytes, and the other cases' thresholds (fresh > 20, savedChars > 1000, …) are measured on one scenario everywhere.

Validation

Windows 11, CODEGRAPH_KERNEL=0 and CODEGRAPH_KERNEL_EXPECT=1, base 0abae11:

test file	fixture line endings	result
main's	LF	fails: call 2 re-sent internal/usecase/payroll/cycle.go lines 31,32,33,34,35 (wasm and kernel)
main's	CRLF	passes 26/26 (the false green this machine saw)
this PR's	CRLF	passes 26/26 (wasm and kernel)
this PR's	LF	passes 26/26 (wasm and kernel)

• The normalisation makes the scenario platform-independent. With it in place but main's strict assertion put back, a CRLF checkout fails the way an LF one does, on lines 31-37 (the 7-line struct).
• The new assertion still catches a real duplicate. With dedupeRange mutated to re-send held runs of 8-23 lines while longer ones still get pointers, the case fails: call 2 re-sent internal/usecase/payroll/cycle.go lines 42-51 (10 lines).
• tsc with the repo's strict options on the test file (temporary tsconfig extending the repo's): clean.
• Full suite on this commit (kernel built, CODEGRAPH_KERNEL_EXPECT=1 npx vitest run --maxWorkers=4 --minWorkers=1): 696 files passed, 6 skipped; 8,347 tests passed, 0 failed, 82 skipped; vitest exit 0.

No CHANGELOG entry: this is a test-only change with nothing user-facing.

Fixes #2756

Notes

• Contributor PR test(explore): a held run shorter than the dedup threshold may come back inside the block around it (#2756) #2758 (@danusha2345) reached the same diagnosis and changes the assertion along the same lines. This was written separately and also normalises the fixture, since the CRLF checkout was what hid the failure on Windows.
• Left out: explore's per-file budget counting \r on CRLF files, and the \r that ends each fenced line there, are product behaviour on Windows checkouts. That's outside this test fix.

🤖 Generated with Claude Code