Story-memory editing and chat layering

Report: https://yumina.io/app/community/thread/73f2ffb9-a1ae-43b2-a954-7dbb48c7e6d4

Behavior

Profile story memory and Settings > AI Configuration story memory/response length keep typing in a local draft. Blur or Enter validates and saves a complete whole number. Escape cancels; an empty or malformed draft restores the existing value. A value outside the allowed range is clamped only when the edit finishes. Preset buttons still apply immediately.

Editing uses plain digits; correctly grouped pasted integers are parsed for the current locale. Invalid suffixes, decimals, exponents and unsafe integers do not silently become truncated token counts. An untouched inherited value is not saved as an explicit choice.

Both profile GET and PATCH responses supply defaultStoryMemory, derived from the same server rollout rule as generation. New accounts in the configured cohort show 16,000. Older accounts and installations without the rollout flag inherit their effective window. Explicit story-memory preferences always take priority. No migration or change to the rollout date is required.

The built-in sandbox Chat now forms a positioned stacking context at z-index 0. This keeps it above preceding creator backgrounds at that same level without lifting it above the creator's higher-level panels. The reported card is “The foundation they ignored, I will rewrite”; its preceding fixed background explained the invisible input while positioned toolbar controls remained visible.

Verification

• Input regression test fails with the original NumberInput and passes with the draft-based implementation.
• Tests cover clearing/backspace, partial typing, Enter, Escape, blank/invalid input, bounds, cap changes during editing, preset changes, and integer grouping in six locales.
• Server tests cover the rollout instant, older/newer accounts, absent/invalid flags and missing account dates.
• Independent review covers draft handling, locale parsing, profile metadata and creator layers.
• The existing back-navigation source audit was updated to the admin page's already-existing search-v3 and filters-v3 keys; application behavior is unchanged by that test maintenance.

The owner approved the interactive preview. Per their browser preference, no browser was opened: the actual reported game still needs a rendered session check. The layer diagnosis is based on the posted screenshot, the actual card root source and CSS paint order; unit tests and a successful build do not replace that visual check.