RegExp ^ and $ under /m recognise only \n — 46 multiline regexes in the cc bundle mis-parse CRLF input

[proggeramlug]

#9218 fixed . to exclude all four LineTerminators. The anchors were not fixed: ^ and $ under /m still recognise only \n.

"one\rtwo".match(/^.*$/gm)      // node ["one","two"]        perry null
"a
b".match(/^.*$/gm)      // node ["a","b"]            perry null
"one\r\ntwo".match(/^.*$/gm)    // node ["one","","two"]     perry ["two"]
"one\r\ntwo".replace(/^/gm,">") // node ">one\r>\n>two"      perry ">one\r\n>two"

Per ES §22.2.2.6, ^/$ in multiline mode match at any LineTerminator: \n, \r, U+2028, U+2029 — the same set . excludes.

Reachability

A tokenizer census (not a grep) of cli_2.1.112.js found 46 /m regexes, including:

/^gitdir:\s*(.+)$/m
/^diff --git /m
/^ID=["']?(\S+?)["']?\s*$/m
/^## /gm
/^#+\s+(.+)$/m
/^[ \t]*<claude-code-hint\s+([^>]*?)\s*\/>[ \t]*$/gm
/[ \t]+$/gm

Any CRLF markdown, git output on a Windows checkout, or /etc/os-release parse silently mis-matches. A cc-level trigger reachable without the network could not be constructed, so this is reachability by static evidence rather than an end-to-end demonstration — but the divergence itself is deterministic and minimal.

Suggested fix

Wherever #9218 established the LineTerminator set for ., reuse it for the multiline anchor translation. The regression test should cover \r, \r\n (which must produce an empty match between the pair), U+2028 and U+2029, in match and replace, since \r\n is the case that most naturally comes out wrong.

Found by a differential stress-test of claude-code under perry.

[proggeramlug]

Closing: fixed in #9427.