# Issue 25157 server proof
Environment: fs-03 Linux, Node26.12.1, isolated worktree .t3 state; vp run dev, Chromium collaborative preview 1280x800; real Codex GPT-6.1-Sol provider via configured hub. No live user data was written.
Before: main53896aea76. Asked for >=100KB fictional manual, no tools. Provider generated several KB before dev watcher restarts interrupted it. Inspector JSON.parse probe captured a 1890-byte full payload parse with stack parseEncodedPayload at ProjectionStore.ts:1286 and turn-item.updated at :2620. This establishes real streaming ingestion, not a latency benchmark. The supplied prior handoff separately observed 9 redundant payload parses.
Red: committed test e08df2759c, command vp test run apps/server/src/orchestration-v2/ProjectionStore.test.ts -t 'indexes streamed payload'. Both128KiB payloads parsed: expected [] but received ['message:streamed-payload-parse','item:streamed-payload-parse']; 1failed46skipped. The test's extra driver field was removed during typecheck in the implementation commit.
After: fd5d54523b. Asked same real provider for1500words about a fictional telescope, no tools. The normal command/event path committed repeated turn-item.updated updates from607bytes to9756bytes, followed by completed message.updated. Inspector probe counted zero matching whole-payload parses in ProjectionStore.apply; see after-probe.json. SQL read-only transcript lists committed sizes/timestamps and verifies all message and turn-item updated_at columns, plus message created_at, match payload JSON (zero mismatches). Browser rendered the full final reply and its final paragraph after completion. Both affected paths are covered at128KiB by the automated regression; the live after reply is9756bytes, not100KiB.
Green: all47 ProjectionStore.test.ts tests passed; check:changed format/lint/knip/graphql/typecheck/tests passed. Both independent GPT-6.1-Sol and Claude Opus5.5 reviews returned MERGE. No rendering changes; visual fixture matrix empty. No end-to-end CPU/latency claim.