Summary

• Stabilize long-session latest-tail anchoring against the target turn end, not only target visibility.
• Keep a static-history fast path when the latest rendered item already fits the viewport and is already near the natural bottom.
• Strengthen the long-session perf E2E guard so a session is not considered usable when the latest content is visible but the viewport is still far above the tail.

Root Cause

The static initial-history path could treat a tall latest model round as successful once any part of that turn was visible. In dense history, that left the final viewport about 1068 px above the physical bottom while still showing latest content, so the old E2E checks did not catch it.

Measurement

Release-fast build, same long-session fixtures, latest upstream baseline 7e69cc5. Candidate values are two runs; baseline is one latest-main run. Negative delta is faster or closer to the tail.

Visual guard results for dense and mixed candidate runs: post-latest loading events 0, blank surface events 0, scroll-jump events 0.

Risk / Impact

• Completed historical sessions whose latest turn is taller than the viewport now bias toward the latest turn tail instead of the middle of the turn. This matches the long-session open expectation, but it is a deliberate UX behavior change for that edge case.
• Static-history fast path skips the stabilization loop only when the latest rendered item fits the viewport and is already within the bottom tolerance. Tall dense items still use stabilization.
• The mixed first-open usable metric showed a small positive delta in two candidate samples, while latest-visible, warm reopen, and rapid-switch median were neutral or better. I do not treat this as a confirmed product regression, but it is listed here because this PR touches session-open timing.

Verification

• pnpm run type-check:web
• pnpm run desktop:build:release-fast
• pnpm run check:repo-hygiene
• pnpm --dir tests/e2e run test:perf with dense fixture perf-anchor-afbff844-dense-a-001
• pnpm --dir tests/e2e run test:perf with mixed fixture perf-anchor-afbff844-mixed-a-001
• Regression RED was verified before the runtime fix: the stricter tail-anchor assertion failed on the dense fixture with latest-main behavior.