Reduce private source model allocation cost

[iperev]

Summary

• Transfer locally owned atomic/layout projections into the private source model instead of copying them twice.
• Preallocate requirement/origin slices from the already-admitted member total; retain reference-edge compaction, the input snapshot, both preflights and all detached accessors.
• Add independent caller/accessor mutation, later-normalization and retained-capacity regression checks, plus allocation benchmarks.
• Prepare compatible release 0.14.20. Public source formats, CLI semantics, resource limits, dependencies and platform support are unchanged.

Proof Boundaries

The paired source-bound experiment compares 111 successful projections, 17 rejected inputs with exact expected diagnostics, and empty/singleton/unequal-group cases. Six samples at each of 64, 512 and 4096 members show approximately 24-29% lower median B/op for this finite family. This is not a universal speedup, peak-RSS claim or proof of a larger safe input domain.

Three isolated mutations are rejected by the permanent tests: wrong member cardinality, removal of reference compaction and an aliased atomic accessor. Full product and provider checks are attached to the exact PR head; review evidence does not replace them.

This closes only the resource-ownership optimization prerequisite. It does not activate source-v2, complete Batch B, replace the deferred human authoring study or close the global roadmap. Future versioned resource-domain restrictions require explicit migration refusal before writes, never silent truncation.

Retrospective

An early copy-only candidate retained excess backing capacity and was rejected before delivery. The first review package also lacked transitive producer and complete benchmark-input evidence; a new frozen package added the actual owners, full source/binary bindings and explicit rejection cases. Original failed and unresolved evidence is preserved. No active skill or provider policy was modified.