Found by the term-fidelity validation matrix on pure relay 10.6.3 (post-#398): low-rate, race-shaped divergences that pass on retry. Evidence: codex resize-mid-stream bundle with renderer cursor off-by-one vs broker ([39,2] vs [40,3] — same signature as the pre-v10 401-cell permission bundle); opencode typing-during-stream where the reconciler fired+repaired live (repair #1). Prime suspect: predictive-echo interaction — upstream erasePredictionRegion early-return persists in relay 10.6.3 (verified during the #398 audit). Leading direction from the fixer investigation: byte LOSS (a dropped erase produces doubled append lines + retained transient rows + row shift). Repro: npm run test:term-fidelity -- --cli=codex (resize workload) / --cli=opencode (typing). Post-#400/#401, every production occurrence self-captures a corpus bundle under userData/term-fidelity-corpus/. Method note: reproduce with PEAR_DIAG_PTY=1 per-chunk diagnostics; do not infer mechanism from aggregate byte counts.
Found by the term-fidelity validation matrix on pure relay 10.6.3 (post-#398): low-rate, race-shaped divergences that pass on retry. Evidence: codex resize-mid-stream bundle with renderer cursor off-by-one vs broker ([39,2] vs [40,3] — same signature as the pre-v10 401-cell permission bundle); opencode typing-during-stream where the reconciler fired+repaired live (repair #1). Prime suspect: predictive-echo interaction — upstream erasePredictionRegion early-return persists in relay 10.6.3 (verified during the #398 audit). Leading direction from the fixer investigation: byte LOSS (a dropped erase produces doubled append lines + retained transient rows + row shift). Repro: npm run test:term-fidelity -- --cli=codex (resize workload) / --cli=opencode (typing). Post-#400/#401, every production occurrence self-captures a corpus bundle under userData/term-fidelity-corpus/. Method note: reproduce with PEAR_DIAG_PTY=1 per-chunk diagnostics; do not infer mechanism from aggregate byte counts.