[LuisAlbertoMK]

What

__tests__/explore-cross-call-dedup.test.ts fails on an LF checkout — i.e. on Linux/macOS CI. It passes on Windows only because CRLF pushes the fixture past the first call's byte budget.

Why it fails

The test asserts that call 2 re-sends no line from call 1. That is stronger than what dedup promises. In dedupeRange, only covered runs of at least EXPLORE_DEDUP.MIN_COVERED_LINES (8) are withheld; shorter runs are emitted deliberately ("cut chunks, not slivers" — an 8-line run is the shortest pointer worth its own ~140 chars).

With LF fixtures the byte budget fits the whole payroll-go fixture into call 1, so call 2 has nothing fresh to send. It falls back to the anti-abandonment restore (suppressedFallback), which puts one withheld file back in full — 5 lines of internal/usecase/payroll/cycle.go — alongside the pointer. CRLF versions of the same 19 fixture files add ~19 bytes each, truncating call 1 and leaving real work for call 2, which hides the problem.

Reproduction on efff7457 with LF fixtures:

AssertionError: call 2 re-sent payroll-go internal/usecase/payroll/cycle.go lines 31,32,33,34,35

What this changes

Test-only. The assertion now matches the contract:

• a served run of at least MIN_COVERED_LINES is never re-sent;
• shorter runs may reappear (they are emitted by design);
• when a call found nothing new (freshLines === 0), exactly one file comes back — the restore, not a dedup leak.

The pointer assertion is unchanged.

Verification

• Pristine efff7457 worktree, 3× LF and 3× CRLF: 26/26 passed in all six runs.
• Mutation check: with the covered.length === 0 guard short-circuited in dedupeRange (dedup effectively off), the suite fails with expected [] to deeply equal [ { start: 100, end: 140 } ] and the missing "Already sent earlier in this conversation" pointer.
• No source changes; nothing else in the suite touched.

[colbymchenry]

Thanks for tracking this down, @LuisAlbertoMK, and for the clear write-up of why only LF checkouts failed. The same fix landed as #2772 shortly after you opened this: it allows a held run shorter than MIN_COVERED_LINES just as you proposed, and also gives the suite's fixture copy LF line endings so every platform tests the same bytes. It's a test-only change, so nothing ships; we're closing this as covered by #2772.