docs(perf): record rejected owned server batch screen

Author: louistrue

scripts/perf/README.md

@@ -843,6 +843,19 @@ it does not establish a performance gain or neutrality. Keep endpoint readiness,
 cache completion and offline witness cost separate, and never extrapolate a
 processing-probe gain to the shipping HTTP artifact.
 [Sanitized screen and limitations](evidence/server-json-roundtrip-4064/README.md).
+### Rejected owned server mesh-batch transfer (#4066)
+
+An owned sink in the canonical processing loop removed the server bridge's deep
+mesh-buffer copy while preserving borrowed callers, retained output, batching,
+progress, styling and cancellation. Exact output and actual cache replay passed,
+but the prespecified HTTP readiness continuation gate failed: the small and MEP
+models were slower in their single pairs, and the largest model's modest time
+improvement accompanied higher sampled RSS. The implementation was not landed
+or expanded to the wider corpus. A source-level copy removal is not an
+end-to-end gain; unbounded downstream ownership and other pipeline work remain
+relevant costs. This screen does not establish precise attribution or a physical
+memory benefit. [Complete verdict, limits and reproducible rejected source](evidence/rejected-owned-batches-2026-09-07/README.md).
+
 ## Current-source native PGO qualification (#4059, not shipped)
 
 Fresh profile training on five public fixtures produced a broad held-out

scripts/perf/evidence/rejected-owned-batches-2026-09-07/README.md

@@ -0,0 +1,55 @@
+# Rejected owned server mesh batches (#4066)
+
+The candidate transferred non-retained mesh batches into the existing server
+channel instead of cloning their owned buffers. It used the same canonical
+processing loop and preserved the borrowed API, retained output, batch/progress
+boundaries, styling and cancellation. It was rejected by the prespecified actual
+HTTP continuation rule. No production implementation or wider corpus run landed.
+
+[screen.json](screen.json) contains every arm of the initial three-model screen,
+its exact gates, separate startup/readiness clocks, sampled RSS, continuation
+rule, source/artifact/compiler identities and validation limits. The small model
+and MEP model were slower in their single pairs; the largest model improved
+modestly while its sampled RSS increased. These observations neither establish
+precise causal contributions nor justify a general claim that ownership transfer
+cannot help. Removing a real copy does not by itself establish an end-to-end win.
+
+All three models passed the declared complete semantic witness, strict raw
+geometry/data-model comparisons, actual hash-only cache replay and cleanup gates.
+No diagnostic mismatch was waived. The primary clock ends when complete SSE and
+data-model receipt both finish; startup is separate. This is HTTP qualification,
+not browser readiness or physical-footprint evidence. Every offline decode and
+retention operation occurred after both timed arms. The isolated verifier reused
+only complete byte-equal inputs with validated witness/version identities;
+current completion and transport fields were independently checked.
+
+The workspace tests and strict workspace/all-target Clippy passed on the exact
+candidate. Downloaded external fixtures were absent: the preserved test evidence
+records ignored tests and fixture-skip messages. Inline/committed-fixture tests
+exercise ownership, both APIs, retained output, styling/progress/bootstrap
+sequences and cancellation. They complement the actual HTTP cases and are not a
+substitute for corpus qualification.
+
+## Reproduce the rejected source
+
+[measured-candidate.patch](measured-candidate.patch) archives the complete local
+candidate, including its tests and Rust API documentation. It uses zero-context
+full-index Git format so patch-context blank lines do not trigger the repository
+whitespace gate. In an isolated checkout at the public base recorded in
+[source-archive.json](source-archive.json), run:
+
+```sh
+git apply --unidiff-zero /path/to/measured-candidate.patch
+```
+
+Temporary-index application reconstructed the **exact measured candidate tree**.
+The public base's Rust/server/Cargo source also matches the measured control.
+The archive records these identities and the patch digest; applying it is not a
+new build or performance measurement. The candidate used the same pinned
+compiler, explicit standard-library build configuration, target and release
+profile as the control, with no PGO inputs.
+
+Raw captures and complete private logs remain retained; repository evidence
+contains sanitized measurements and their hash references, not those large
+payloads. Exact duplicate evidence files share APFS extents while retaining
+separate paths and inodes. No unique raw evidence was removed.