[zakazaka95]

Continues #4020 following the invitation to resubmit, and addresses #3914 against current main (301cbb99).

Long requests can lose the beginning of a line from the diagnostic window, causing the reported column to restart inside the excerpt. Track discarded rune starts and trim at UTF-8 boundaries so the reported position remains absolute while the caret stays aligned with the displayed text. This retains current main’s JSON-v2 error handling and comma adjustment.

The three chunked cases now have literal expected JSON responses in a separate TestHandleParseErrorChunked, as requested in the earlier review. The long formatting expression is split as requested too.

Validation

Linux validation and benchmark run:

• Full go test -short ./jsonrpc and go vet ./jsonrpc pass.
• Both repository lint configurations pass for JSON-RPC, including strict changed-line checks; gofmt/gofumpt pass.
• Restoring main’s formatter, removing absolute-prefix accounting, or using the absolute column for the displayed caret makes all three targeted cases fail. Restoring the fix passes again.
• Dependencies and lockfiles are unchanged. No full-node or whole-repository test claim.

Requested benchmarks

Five runs per case, 2 seconds each, CPU=1 and GOMAXPROCS=1, on the same Linux runner. Values below are medians against 301cbb99.

Benchmark	Main ns/op	Fix ns/op	Change	B/op main → fix	Allocs/op main → fix
BenchmarkHandle	3,586	3,596	+0.28%	1,036 → 1,039	18 → 18
LargeRequest/1MB	3,365,575	3,589,162	+6.64%	7,358,860 → 7,358,868	52 → 52
LargeRequest/10MB	26,696,681	27,707,726	+3.79%	65,035,847 → 65,035,849	58 → 58

These benchmarks use ASCII request bodies. The regression cases also cover mixed-width UTF-8 and a discarded newline. The extra prefix scan costs time on large requests; an ASCII-word fast path reduces that cost.