The scanner's count-based cost guards measure scanAllPatterns alone, so the rest of the per-line scan is outside every cost guard. probeWork (internal/adapter/scanner/adjacency_test.go) tallies only the bytes handed to the adjacency probes and the junction probe inside scanAllPatterns; the wall-clock guards it replaced timed scanLine, which is the whole of scanText. Uncovered stages: (1) each pattern's Skip and SkipAt callbacks, run per match against the matched text and the whole line; (2) the identity matchers (matchers.findings), run over the raw line and again over the decoded copy; (3) the percent-decode pre-pass (decodedLineFindings: percentDecodeBounded, the second scanAllPatterns run over the decoded copy, mapDecodedSpan); (4) the top-level cp.Re.FindAllStringIndex pass per pattern, which is RE2-linear but untallied. A regression that makes any of these superlinear passes every guard in the package. Counting them at the scanLine level needs a seam in scanText, which builds its probes internally and constructs the identity matchers from a fixed type with no hook for a counting stand-in, so it is a production change rather than a test-only one; hence captured, not added, in the gallop lane.

The scanner's count-based cost guards measure scanAllPatterns alone, so the rest of the per-line scan is outside every cost guard. probeWork (internal/adapter/scanner/adjacency_test.go) tallies only the bytes handed to the adjacency probes and the junction probe inside scanAllPatterns; the wall-clock guards it replaced timed scanLine, which is the whole of scanText. Uncovered stages: (1) each pattern's Skip and SkipAt callbacks, run per match against the matched text and the whole line; (2) the identity matchers (matchers.findings), run over the raw line and again over the decoded copy; (3) the percent-decode pre-pass (decodedLineFindings: percentDecodeBounded, the second scanAllPatterns run over the decoded copy, mapDecodedSpan); (4) the top-level cp.Re.FindAllStringIndex pass per pattern, which is RE2-linear but untallied. A regression that makes any of these superlinear passes every guard in the package. Counting them at the scanLine level needs a seam in scanText, which builds its probes internally and constructs the identity matchers from a fixed type with no hook for a counting stand-in, so it is a production change rather than a test-only one; hence captured, not added, in the gallop lane.

Grounds

• pursued: a per-line stage whose work outgrows the line fails TestScanLineWorkIsLinear on some shape; a superlinear helper that charges nothing to the meter, or a stage no fixture exercises densely, would show it wrong.