fix(neopi): preserve reader position through streamed completion #168

Observed behavior

During private native-workspace verification for #167, a reader who scrolled to the first message while an assistant response was visibly streaming was pulled near the end after completion, without further input. Stable finding: CORE-1 — observed once, unreproduced in the corrected diagnostic, cause unproven, not fixed.

The failing clean build used NeoCode 39579c45638afe36858fe2fea2d1ba21443e1848, with bundled NeoPi dcc5fed8eb9a9b4dcac3a15eec106c27b540720c / SHA-256 134a49efefd126ede6e1299daac4305ed5f4d39790743bf151dd59ae50d4e22d. The main window was 1600×1000, with four preceding conversation rows. “395” identifies a source revision, not a 395px viewport.

Terra HIGH captured an incomplete second item while Working (16-partial-stream), then the original DRAFT_SHARED_ALPHA row and scrollbar thumb at the top while Working (17-reader-up). With no subsequent input, 18-reader-pulled shows Ready, items 3–10 and the thumb near the bottom. Main and independent Opus 5.5 HIGH both viewed those captures; the no-input interval is recorded by Terra.

Exact submitted prompt:

Explain ten practical ways cities can improve public spaces. Give each a short paragraph, then end with STREAM_FINISHED. Do not use tools.

Diagnostic evidence and limits

• Earlier private instrumented trials did not reproduce the pull, but used an 800px-wide window and additional preceding history. They are not equivalent reproductions of the original layout. One earlier trial also wheeled before a new assistant body existed; Working alone was insufficient.
• The corrected original-branch trial restored the four preceding rows using native Session tree navigation to assistant entry d2e5290f, preserved later branches, used 1600×1000 and the exact prompt, and wheeled only after new incomplete body text appeared. At 13:16 UTC on 2026-10-29, the trace records the wheel and its immediate cancellation write, then live-row growth from 292.875 to 922.25 pixels for 17.6 seconds. The viewport remained at the top through Ready; no later follow, maintain, adjust or padding write explained the historical failure. This is a strong non-reproduction, not a fix or a causal account.
• Instrumented source was private diagnostic patch SHA-256 7af96e34cac0fe867559a704b597779c59b1fffbed82264814fde94d18bf3f1c over the same clean source/native pin. Product probes were removed before the next shipping build. Instrumentation may perturb scheduling; its effect is unmeasured.
• Current refactor(neopi): make native sessions the client authority #167 head 8ac8171d095a08735670d087891cdc1ec94cf32a repairs browser-sheet layering and contains no scroll change. Independent review retains CORE-1 as unresolved.

Private reports, screenshots and timestamped write logs are retained by the maintainer. They are not publicly uploaded; these filenames identify evidence, not public links. No user journal was rewritten and no native cause has been established.

Desired behavior and acceptance

• Deliberate upward navigation during a streamed response remains respected through later row growth, durable replacement and completion.
• A failing interval identifies the first unexpected scroll-top change and its initiating programmatic write or layout event before choosing a repair.
• Any repair is verified through the actual native transcript, including continuation after the reader deliberately leaves the end. Passing unrelated commands or later non-reproductions do not establish a fix.

Do not apply speculative follow/MVCP changes and call this resolved. The two commissioned attempts are complete and recorded in the bounded-attempt evidence comment: both were non-reproductions, not fixes. No further scroll experiments are commissioned.

Native core accepted with this issue open

On 2026-10-29, Saturn reported the owner's delegated decision: “use your best judgment, I trust you.” Saturn selected (a): accept the native core with CORE-1 tracked and open. This clears the core gate for the staged R1 design build and SQL cutover, with a Terra walkthrough followed by incremental independent Opus review per stage.

This decision does not change the evidence: CORE-1 was observed once, remains unreproduced at the target failure interval, its cause is unproven, and it is not fixed. The acceptance checkboxes above remain unchecked. Mini-player return behavior still awaits the owner's intent and is not decided here.

Earlier coordination instructions — retained as history

Also file CORE-1 as a tracked issue now with its evidence and both diagnostic runs, so whichever way she decides there's a record.

Saturn: one clarification on (b). It's provisional, the default hold while she's away from work, not a closed decision. The trace doesn't foreclose anything: if she chooses (a) when she reads the morning report, the gate opens with CORE-1 tracked, and R1/SQL proceed. Don't treat (b) as her ruling.