opencode typing-during-stream: genuine content divergence (reconciler repaired) — root cause unknown

[khaliqgant]

The last unexplained rendering event from the term-fidelity program (2026-07-17). During the v10 validation matrix, opencode's typing-during-stream workload produced a GENUINE content divergence that the reconciler detected and repaired live (repair #1, bundle under the preserved v10-residual-bundles + reproducible via npm run test:term-fidelity -- --cli=opencode, intermittent). NOT the #403 viewport-pinning bug (that was content-perfect); this one had real differing cells. Tooling ready: the raw-stream discriminator merged in #416 auto-classifies any recurrence as RENDERER-INTERNAL vs DELIVERY-LOSS, and #400/#401 corpus+telemetry capture production occurrences automatically. Method (from the #403 investigation): PEAR_DIAG_PTY=1 byte-level capture, evidence before mechanism claims, suspects include predictive-echo interaction during typing (upstream erasePredictionRegion flaw, relay#1326/#1327 adjacent).

[khaliqgant]

Closing as not currently reproducible on main — a verified no-longer-reproduces result (evidence-first, bounded, denominator logged).

Bounded hunt (N=28, hardened harness, opencode typing-during-stream): 28/28 reached readiness (vs ~1/8 pre-harness-fix), 27 checkpoint-reached runs, 0 divergences, 0 reconciler telemetry — the #416 raw-stream discriminator was armed on every run and never fired. Total across this hunt + prior clean runs: ~37 clean typing-during-stream observations, 0 divergences.

Along the way we root-caused (and fixed, #418) a separate harness readiness flake — the broker's spawn-time task injection racing opencode's slow (~13s) OpenTUI boot — which had made the divergence look more common than it is; opencode's rendering + model backend were both proven healthy standalone. The db-contention hypothesis was refuted by a direct lsof test.

Most plausible prior mitigation: #398 (v9-JS/v10-daemon skew), #408 (pty-dedup re-key), #412 (viewport-pin). The #416 discriminator stays taped on in the matrix, so any future firing is auto-classified (RENDERER-INTERNAL vs DELIVERY-LOSS) — reopen if it ever fires. No mechanism was ever asserted; a direct probe never confirmed one because the phenomenon never fired.