[jaredLunde]

Fixes D3. The TypeScript client re-matched long lines with a JavaScript RegExp, so Rust-syntax patterns ((?i), (?P<x>..)) failed. No client re-matches now: rg --column reports the byte column of the first match, and every client turns it into a character offset the same way (UTF-8 encoding of the decoded line, CRs not counted; a column inside a character means no column).

Spec decisions (Clipping, Parsing, Engines differ):

• rg has no flag that reports the match end alongside the line, so the window is placed by the match start only: start = max(0, cs - 250); start 0 is the head clip; a window past the end moves back to the last 500 characters. Previously it used the match width.
• GNU grep and busybox report no column, so those rungs clip every long line at its head (they used to re-match client-side).
• Scenarios: files.search.long-lines-are-clipped now requires rg with the new window numbers; a grep-rung twin pins the head clip. The differential oracle and the ShellFs property oracle follow.

Removes the D3 xfails and tolerance. Matrix: all search scenarios pass on local/alpine/debian-runas/readonly/tinydisk for rust, ts, python, python-async. The sandbox-mcp driver fails only the D5 scenario (next fix). Differential: 6/6 with the four client drivers (the mcp driver diverges on a plain mkdir in the differential, unrelated to this change).

Stacked on #594 (D2) and carries #590; retarget once they land.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HJRTKuPpNNqCLEdyUwZj64

[jaredLunde]

Landed on main as part of #600: the three D fixes were stacked, and #600's squash carried this PR's commits.