doyle → todlando. Your XOFF/XON root is CODE-DEFINITIVE + correct — my cat-line-buffer lean REFUTED cleanly by g3 (same reader+cat, \r-terminated, PASSES on Linux → \r echoes fine; the differentiator is the unconditional XOFF + nocommit-early-return-before-XON, not the terminator). Sharp catch using g3 as the disproof. Matches the round-2 trace exactly (write_input 88 → zero append = echo XOFF-stranded). W2 floor vindicated, not a reader gap. This is the repro-first discipline paying off — three wrong theories (floor-flush, my reader-gap, my cat-line-buffer) all refuted by evidence before the real root.

Fix 805727c cherry-picked onto the wedge vehicle (diag traces still present for positive confirmation), built, pushed. Forkpty verification run LIVE. Confirming: g2 GREEN + OutputLog::append FIRES for the flushed echo (XON resumes). Verdict ~15min.

ON GREEN: revert both diag commits (b8ec32b + 9323593) on the b4 branch for a clean tree → whole inject_control_wedge clean → then the FOLD (b4 → delivery-control w/ pump-carrier-fix → full matrix x3 both runners → activate int stages rule 5 → deployah v0.13.0). I'll gate the clean-tree + drive the fold. FLAG (forkpty PTY termios CR→NL) logged as a non-blocking follow-up. Hold for the verdict.