Allocation-intent hint: tell the RTL growth policy how a buffer will be used

Track O (optimization; heap allocator + a thin frontend flag) — lands under A's gate (self-host byte-identical). Deferred / earn-it-first: no live profile is pulling this. Filed so the design is on record, not to schedule it.

Problem

The allocator cannot see intent. A one-shot s := s + c (a directory path, a known-length transform) and an unknown-length accumulation in a loop hit the SAME growth path, so a single global growth policy is wrong for both. The intent is legible at the CALL SITE / in the AST, not at runtime.

Shape (small — two flags, not a predictive model)

Thread a tiny intent enum from the frontend through the managed-string / dynarray alloc lowering into the RTL growth decision:

• exact / compact — final size is known or the value is done being built → allocate exact, no slack, trim on finalize. (Most s + c one-shots, upper/ lower, path building.)
• growing — accumulation whose length is unknown up front → amortised growth (1.5x = n + n>>1, sub-golden so freed blocks stay reusable; see the reuse note in [[idea-adaptive-heap-growth]]), page/size-class-rounded, ratio capped by a per-target slab so 1 GB + 1 B never doubles to 2 GB.

The free 80% (do this part first if anything)

A function's Result string being built is ALREADY special (the return / NRVO path), so the RTL can auto-apply growing slack + trim-at-return to any Result-typed managed string with ZERO frontend work. That captures the common "build a string and return it" case. The explicit growing/exact flag is only needed for non-Result locals accumulated in a loop.

Relationship

This is the middle slice between the concrete [[perf-nilpy-remaining-perbyte-string-builders]] (fix known sites by hand, do now) and the research [[idea-adaptive-heap-growth]] (which supersedes this). Do the smallest slice a profile actually demands; never top-down.

Gate

make test + self-host byte-identical (touches shared RTL / alloc lowering) + cross where a target's growth policy differs (ESP wants a smaller cap). Land the Result-auto-slack part alone first; the explicit flag is a follow-on.