Problem

The session mirror produces word-merged text like I'llreadthecodebase because the Ink TUI renders adjacent characters without column gaps. The cursor-tracking parser in mirror_strip_ansi() correctly sees no gap but the result is unreadable prose.

Issue #29 accepts this limitation for the initial fix (plan B+D+F). This issue tracks the harder problem of splitting merged words in the live mirror.

Approach: Hybrid strip + selective wordninja

1. Replace mirror_strip_ansi() with the simpler strip_ansi() (regex replace with space + collapse) for mirror output
2. For lines that look like prose (>40 chars, no code indicators like /, \, ::, {, }), run wordninja.split() and rejoin
3. Leave code-like lines (paths, commands, SQL) untouched

Performance

• wordninja.split() is ~2ms per call on typical lines
• With rate-limited mirror writes (200-500ms buffer from Session mirror: garbage not filtered + word-merging/letter drops #29), the number of lines processed is already reduced
• Only prose lines get split — code/paths are skipped
• Acceptable for a log file that's consumed by tail -f

Depends on

• Session mirror: garbage not filtered + word-merging/letter drops #29 — shared filter module + rate-limited writes (implement first, then evaluate if word-merging is still a problem after rate-limiting reduces mid-repaint artifacts)

Alternative

If wordninja is too aggressive (splits identifiers, URLs), consider a lighter approach: insert spaces before capital letters in runs of 30+ lowercase chars. Catches readthecodebase → read the codebase but won't help with Doyouwanttoproceed.