New procedural content keeps every existing seed's output only if it is generated last, from a random stream of its own

Fact

A generator that draws everything from one seeded stream produces a different world for every seed the moment a new step consumes a draw, or merely changes whether an earlier step accepts a candidate (a new "keep clear of X" check in a rejection loop). Everything drawn after that point shifts. Adding the new feature at the end of generation, from its own stream (new Rng(seed ^ CONST)), and making it edit the finished result (clearing what is in its way) leaves every earlier step bit-for-bit identical.

Why it matters

Tests pinned to seeds silently become tests of different content. Statistical assertions that only ever passed on a lucky seed ("most agents do X on seed N") start failing, and the failure points at code the change never touched. Agent simulations sharing one stream make it worse: a single entity nudged at t=0 reorders every later draw, so the whole run diverges even when the new content is far from anything being measured.

How to apply

• Put the new step after all existing ones; give it its own derived seed. Do not add constraints to earlier steps' accept/reject loops; let the new step remove or move what conflicts instead.
• Verify by diffing old and new output with the new step switched off: they must match exactly.
• When a seed-pinned behavioural test still fails, measure the property across many seeds on the old code before blaming the change: if the old code already fails on a large share of them, the test is seed-lucky, and that is a separate bug.