Read this first: what belongs in Optimization

[GeneralPawz]

This category is for making existing, working behavior faster or lighter.
It is not for asking the kernel to do something new.

What belongs here

• "This operation is O(n^2) and could be O(n log n)"
• "This allocates per call and could reuse a buffer"
• "This build pulls in packages it never uses"

What does not

Instead you want	Go here
A capability that does not exist yet	Features
An API that works but is awkward to call	Papercut issue
Something that behaves incorrectly	Bug issue

Optimization vs papercut: an optimization costs you milliseconds; a
papercut costs you keystrokes. A rename is a papercut. A cache is an
optimization.

What a good proposal contains

1. A measurement. Same harness, pinned toolchain, repeated medians. An
   intuition that something "should be slow" is not evidence.
2. What it trades against. Optimization competes with capability work.
3. What stays true. An optimization that changes results is not an
   optimization -- it is a behavior change, and needs a different conversation.

Please read before proposing

Broad performance work is deliberately parked
(ADR 0013).
The kernel is being built for capability first and shaped for speed later, so
that optimization stays possible -- layering, provider seams, measured
oracles -- without trading capability for benchmarks now.

Benchmarks currently exist for axiolid-reference and
axiolid-mesh-boolean-boolmesh only. There are no broad performance claims to
defend, and none should be made without same-harness evidence.

That does not make this category pointless: proposals here get recorded,
argued, and either accepted with a measurement or parked with a written
unblock condition. Nothing is silently dropped.