fix(discover): count every line of a multi-line call - #5024

Summary

rtk discover and rtk session read a multi-line call as one command and classified only its first line. git status\ngit log counted one command, and cd x\ngit status offered no saving at all, although the rewrite acts on each line.

The report now reads a block as the rewrite does. Both go through one entry check, read_line, and one block_lines: a line is a command, a trailing operator joins the next line, and a blank or comment-only line is no command. A block the rewrite refuses to read line by line, such as a loop, stays one text as before. Each command of a call is read once, for its share of the output and for its row.

Built on #4401. It uses that stack's lexer.

Behaviour changes

• total_commands and session adoption now count every line, including lines rtk doesn't handle, so a multi-line setup script lowers a session's adoption.
• A call has one measured output. It is now shared by the commands that may have written it, in proportion to their category averages, and marked estimated when more than one shares it. A command alone takes it whole, so cargo test\necho done keeps its measurement. Both stages of a pipeline share it. A command that writes next to nothing (echo, cd) takes none, one that sends both its streams elsewhere (cargo test > test.log 2>&1) takes zero, and any other command with no row, as in git status && ./build.sh, takes a share no row shows. This also changes ; and &&: develop gives the number to the last command, and drops it if that is an echo.
• A line left inside a backquote, or a heredoc with an fd number in front (cat 0<<EOF), now refuses its block in rewrite_command. The hook's answer doesn't change: it already leaves any substitution or heredoc alone.

Known limits

• A CRLF block stays one text in the report, while the rewrite reads it line by line.
• A quoted newline beside a real one, such as a multi-line commit message followed by another line, stays one text, because the rewrite refuses that block.
• A comment-only block still counts as one command.
• A redirect written on a subshell or a group ((cargo test) > test.log 2>&1) is read as its own part, so the command inside keeps a share of what came back.
• A line beside a command substitution is listed as reached when the rewrite alone would take it, although the hook defers the whole call; a single-line chain already reads this way.
• A command on the quiet list can still print a lot (set alone, cp -v); its output then goes to the commands around it. One behind a wrapper (command -v, sudo mkdir) is read as a writer.
• Redirects are read one command at a time, so an exec > log 2>&1 does not reach the commands after it.
• Without a hook log row, each supported line re-runs the whole-block permission check, so the estimate is quadratic on very long blocks: about 7 s for 1,000 lines on a debug build.

Test plan

• cargo fmt --all, cargo clippy --all-targets, cargo test --all, and both with --no-default-features
• Each guard fails a test when removed
• 49 multi-line inputs compared before and after: no rewrite moved

🤖 Generated with Claude Code