Intermittent opencode reconciler-repair in full term-fidelity matrix (delivery-clean, renderer-internal transient divergence on alt-screen TUI) #415

Split out from the #403 fixer work (the codex resize vector is fixed in #412). This is the separate opencode vector the coordinator flagged (typing-during-stream, "reconciler fired + repaired").

Reproduction

• Does NOT reproduce as an isolated workload: npm run test:term-fidelity -- --cli=opencode with TERM_FIDELITY_WORKLOAD=typing-during-stream was 12/12 clean. It needs the accumulated full-matrix context (prior workloads / warmed state).
• Reproduces intermittently on the full opencode matrix (--cli=opencode, all 6 workloads): ~2 in 10 runs, observed firing on both typing-during-stream and resize-mid-stream.
• Failure mode is a reconciler repair (telemetry-only): [terminal] viewport diverged from broker screen; repainted from snapshot (repair #1). The matrix fails on any repair by design (to surface creation vectors); the grid self-heals.

Byte evidence (captured resize-mid-stream instance)

Using a raw-stream discriminator (tap onPtyChunk, replay the exact delivered bytes through a clean @xterm/headless at broker dims, diff vs broker):

rawBytes=621184 bufferType=alternate rawVsBroker=0 rawVsRenderer=0 verdict=RENDERER-INTERNAL
telemetry: [resize-mid-stream] viewport diverged from broker screen; repainted from snapshot (repair #1)

• rawVsBroker=0 — the full delivered byte stream, replayed through a clean emulator, equals the broker screen exactly. Delivery is byte-clean end to end.
• bufferType=alternate — opencode runs on the alternate screen.

What this EXCLUDES

• Byte loss / dropped erase / delivery corruption — rawVsBroker=0.
• Predictive echo / erasePredictionRegion (relay#1326) — opencode is alt-screen, and the echo-router kills predictions on alt-screen (onUserInput → if (isAltScreen) killPredictions()). Predictions never engage here.
• The Residual intermittent terminal divergence on v10 (resize/typing during streaming; predictive-echo suspect) #403/fix(terminal): keep the viewport pinned across resize/redraw reflows (#403) #412 viewport-scroll vector — alt-screen has no scrollback to strand into; #412 is a no-op for opencode.

Assessment (narrowing, not a pinned fix)

The transient divergence is renderer-internal: the live xterm's incremental rendering drifted from a faithful batch render of the same (clean) bytes, and the reconciler correctly repaired it from the broker snapshot. This is the AGENTS.md "diff-painting TUI preserves a one-time grid divergence" class the reconciler backstop was built for.

• resize-mid-stream: consistent with xterm reflowing during the resize storm differently than the broker PTY emulator, in the transient window before opencode's post-SIGWINCH repaint. opencode's diff-paint then skips cells it believes unchanged, so the divergence persists until the reconciler repairs. Plausibly inherent (the reconciler may be the correct/only handling), which would make the matrix's "fail on any repair" too strict for this class — worth a policy discussion.
• typing-during-stream (coordinator's cited case): NOT separately byte-captured yet. A pure-content live-vs-batch divergence would be more surprising (xterm's streaming parser is batch-equivalent for content), so this one may have a distinct creation vector (e.g. an input-echo / coalescing ordering interaction) and deserves its own capture.

Next steps

• Land the raw-stream discriminator + scroll-probe + telemetry-artifact byte capture into the term-fidelity harness (in the accompanying PR) so any future opencode repair auto-classifies DELIVERY-LOSS vs RENDERER-INTERNAL and dumps raw.bin for the typing case.
• Capture a typing-during-stream repair with the discriminator; if rawVsBroker=0 there too, investigate the renderer write-path ordering under concurrent input+stream; if it's the inherent reflow class, decide whether the matrix should tolerate a bounded reconciler repair rate for alt-screen TUIs.

Isolated brokers only; port 3889 untouched; all work on a git worktree.