perf(memmove): record 4096-byte copy-codegen gate

bd-2g7oyh.391 no-code rejected on RCH vmi1227854.

Broad route: FrankenLibC memmove_4096 p50/mean 35.831/37.441 ns vs host 27.725/29.024 ns.

Focused gate: FrankenLibC Criterion [33.701 ns 34.193 ns 34.727 ns], p50/mean 34.521/34.938 ns; host [26.673 ns 27.185 ns 27.776 ns], p50/mean 27.575/31.903 ns.

No source change: prior copy-panel families rejected; next route is generated/disassembly-backed copy lowering. Score 0.0.

Author: Dicklesworthstone

.beads/issues.jsonl

@@ -6540,3 +6540,4 @@
 {"id":"frankentorch-zz6","title":"Implement tensor creation functions: eye, arange, linspace, logspace","description":"Add torch.eye, torch.arange, torch.linspace, torch.logspace to ft-api. These are used everywhere in model code and tests. All are non-differentiable leaf tensors.","status":"closed","priority":1,"issue_type":"feature","created_at":"2026-03-27T18:14:34.826544482Z","created_by":"ubuntu","updated_at":"2026-03-27T18:15:06.321201936Z","closed_at":"2026-03-27T18:15:06.320936834Z","close_reason":"Already implemented: eye, arange, linspace, logspace all present in ft-api with tests.","source_repo":".","compaction_level":0,"original_size":0}
 {"id":"bd-2g7oyh.389","title":"[perf][no-gaps] strncmp_256_equal focused codegen gate","description":"Pass 90 target from the post-bd-2g7oyh.388 broad RCH non-math routing profile on vmi1227854. Profile evidence: strncmp_256_equal FrankenLibC p50 6.842 ns, mean 9.196 ns vs host glibc p50 5.350 ns, mean 6.464 ns. Prior bd-2g7oyh.65 strncmp SIMD work is closed and already landed; a keepable edit must be structurally different from prior 32-lane equality panels and SWAR strcmp/strncmp follow-ups, preferably a codegen/load-shape primitive with exact first-difference/NUL ordering preserved. Scope: one focused same-worker baseline before any edit; crates/frankenlibc-core/src/string/str.rs only unless proof requires ABI diff test wiring. Required proof: focused RCH baseline, strncmp scalar-reference/golden sha256 and ABI/glibc differential, rustfmt/git diff check, same-worker RCH post, Score>=2.0 else restore/no-code closeout.","status":"closed","priority":1,"issue_type":"task","assignee":"BoldFalcon","created_at":"2026-06-13T19:04:12.225418399Z","created_by":"BoldFalcon","updated_at":"2026-06-13T21:22:37.460687900Z","closed_at":"2026-06-13T21:22:37.460687900Z","close_reason":"KEPT: exact-256 safe-SIMD equality certificate for strncmp. Same-worker RCH vmi1227854 baseline FrankenLibC p50/mean 5.756/7.300 ns vs host 3.867/5.056; post FrankenLibC p50/mean 4.052/5.185 ns vs host 4.501/5.846. Behavior proof passed: core strncmp 2/2, property/golden strncmp 3/3 with golden SHA 99a3358be31072baca18340daceec13300282aa57b2a1b7406d6817396edb326, ABI diff_strncmp_cases 1/1, cargo check -p frankenlibc-core --lib passed. Touched-file rustfmt and git diff --check passed. Strict clippy remains blocked only by existing unrelated lint debt outside the touched str.rs hunk. Score 10.0.","source_repo":"frankenlibc","source_repo_path":"/data/projects/frankenlibc","compaction_level":0,"original_size":0,"labels":["no-gaps","perf","string","strncmp"],"dependencies":[{"issue_id":"bd-2g7oyh.389","depends_on_id":"bd-2g7oyh","type":"parent-child","created_at":"2026-06-13T19:04:12.225418399Z","created_by":"BoldFalcon","metadata":"{}","thread_id":""}]}
 {"id":"bd-2g7oyh.390","title":"[perf] malloc_free_64 focused allocator gate","description":"Screen the post-bd-2g7oyh.389 allocator residual with a same-worker RCH focused baseline before any source edit. Only proceed to a structural slab/LIFO or intrusive-list allocator primitive if the focused gap is material; do not repeat exact-size/cache shortcuts, Trace lifecycle gates, fixed magazine/plain storage swaps, certificate/log micro-specialization, hot-slot metadata tweaks, lazy-accounting retunes, or bitmap-only hot-slot representation changes.","status":"closed","priority":1,"issue_type":"task","assignee":"BoldFalcon","created_at":"2026-06-13T21:42:29.393197950Z","created_by":"BoldFalcon","updated_at":"2026-06-13T21:42:29.393197950Z","closed_at":"2026-06-13T21:42:29.393197950Z","close_reason":"No-code rejected: focused RCH vmi1227854 malloc_free_64 gate collapsed to FrankenLibC p50/mean 4.989/6.027 ns vs host 4.238/11.317 ns; p50 absolute gap 0.751 ns and mean favored FrankenLibC, below allocator edit gate.","labels":["perf","no-gaps","malloc","allocator"],"parent_id":"bd-2g7oyh"}
+{"id":"bd-2g7oyh.391","title":"[perf] memmove_4096 focused copy-codegen gate","description":"Screen the post-bd-2g7oyh.389 memmove_4096 residual with a same-worker RCH focused baseline. Do not edit unless the focused gap is material and the candidate is a generated/disassembly-backed codegen, alignment, or no-overlap primitive, not another exact-size safe-SIMD copy panel, general portable-SIMD copy loop, inline-only hint, or related memcpy branchback.","status":"closed","priority":1,"issue_type":"task","assignee":"BoldFalcon","created_at":"2026-06-13T21:49:48.046730180Z","created_by":"BoldFalcon","updated_at":"2026-06-13T21:49:48.046730180Z","closed_at":"2026-06-13T21:49:48.046730180Z","close_reason":"No-code rejected: focused RCH vmi1227854 memmove_4096 reproduced p50 gap but prior copy-panel families rule out a repeat source lever. Baseline FrankenLibC p50/mean 34.521/34.938 ns vs host 27.575/31.903 ns. Next route is generated/disassembly-backed copy lowering before any source edit.","labels":["perf","no-gaps","string","memmove"],"parent_id":"bd-2g7oyh"}

.skill-loop-progress.md

@@ -911,3 +911,12 @@
 - Verdict: NO-CODE REJECTED, Score `0.0`. The focused p50 gap collapsed to `1.177x` with only `0.751 ns` absolute gap, and mean reversed in FrankenLibC's favor due host tail. No allocator source was touched; allocation order, LIFO/tie-breaking, lifecycle logs, counters, FP, and RNG are unchanged by construction.
 - Artifact: `tests/artifacts/perf/bd-2g7oyh.390-malloc-free-64-focused.md`.
 - Next route: continue from the current broad table with a different focused residual. Only return to allocator with a material focused same-worker gap and a true slab/LIFO or intrusive-list primitive, not another hot-slot metadata or observability micro-lever.
+
+### Pass 92 - Memmove4096 Focused Copy-codegen Gate - 2026-06-13T21:49:48Z
+- Target: `bd-2g7oyh.391` / `glibc_baseline_memmove_4096`, selected from the post-`.389` broad non-math RCH profile after the allocator focused gate collapsed.
+- Broad route basis: `memmove_4096` FrankenLibC p50/mean `35.831/37.441 ns` vs host `27.725/29.024 ns`.
+- Route constraint: prior no-ship families include exact 4096 safe-SIMD copy panels (`bd-2g7oyh.346`), general portable-SIMD copy panels (`bd-2g7oyh.304`), inline-only `memmove`, and related exact 4096 `memcpy` copy-panel/full-slice branchbacks. A source edit requires a materially different codegen/alignment/no-overlap primitive.
+- Focused RCH baseline: `RCH_BUILD_SLOTS=1` on `vmi1227854` showed FrankenLibC Criterion `[33.701 ns 34.193 ns 34.727 ns]`, p50/mean `34.521/34.938 ns` vs host `[26.673 ns 27.185 ns 27.776 ns]`, p50/mean `27.575/31.903 ns`.
+- Verdict: NO-CODE REJECTED, Score `0.0`. The focused p50 gap reproduced (`1.252x`, `6.946 ns`) but the only admissible next step is generated/disassembly-backed copy lowering, not another manual safe-SIMD panel loop. No source was touched; prefix-copy length, bounded `n`, overlap behavior, destination suffix preservation, return value, FP, and RNG are unchanged by construction.
+- Artifact: `tests/artifacts/perf/bd-2g7oyh.391-memmove-4096-focused.md`.
+- Next route: build or reuse an RCH-compatible assembly/codegen extraction artifact for safe string/memory kernels, then return to `memmove_4096` with alignment-aware safe-slice dispatch, array-reference copy lowering with assembly proof, or ABI-level raw-pointer classification that preserves the safe core contract.

tests/artifacts/perf/bd-2g7oyh.391-memmove-4096-focused.md

@@ -0,0 +1,72 @@
+# bd-2g7oyh.391 memmove_4096 focused copy-codegen gate
+
+Date: 2026-06-13
+Agent: BoldFalcon
+Status: NO-CODE REJECTED
+
+## Target
+
+Post-`bd-2g7oyh.389` broad non-math profile on RCH `vmi1227854` showed a
+fresh `memmove_4096` residual:
+
+| row | FrankenLibC p50 ns | host p50 ns | FrankenLibC mean ns | host mean ns |
+| --- | ---: | ---: | ---: | ---: |
+| broad `memmove_4096` | 35.831 | 27.725 | 37.441 | 29.024 |
+
+Prior no-ship families are explicit:
+
+- exact 4096 safe-SIMD copy panels under `bd-2g7oyh.346`;
+- general safe portable-SIMD copy panels under `bd-2g7oyh.304`;
+- inline-only `memmove` hinting under `bd-2g7oyh.157`;
+- related exact 4096 `memcpy` copy-panel/full-slice branchbacks.
+
+Any source edit therefore required a materially different codegen, alignment, or
+no-overlap primitive. No source edit was attempted in this pass.
+
+## Focused RCH Gate
+
+Command:
+
+```text
+RCH_BUILD_SLOTS=1 RCH_WORKER=vmi1227854 RCH_PREFERRED_WORKER=vmi1227854 \
+RCH_REQUIRE_REMOTE=1 RCH_VISIBILITY=summary RCH_QUEUE_WHEN_BUSY=1 \
+RCH_DAEMON_WAIT_RESPONSE_TIMEOUT_SECS=1200 rch exec -- env \
+AGENT_NAME=BoldFalcon FRANKENLIBC_BENCH_PIN=1 CARGO_BUILD_JOBS=1 \
+CARGO_TARGET_DIR=/data/tmp/frankenlibc-bd391-memmove-focused-baseline-target-20260613T2145 \
+cargo bench -j 1 -p frankenlibc-bench --bench glibc_baseline_bench -- \
+glibc_baseline_memmove_4096 --noplot --sample-size 50 --warm-up-time 1 \
+--measurement-time 3
+```
+
+RCH selected `vmi1227854`.
+
+| impl | Criterion interval | p50 ns | mean ns | p95 ns | p99 ns |
+| --- | --- | ---: | ---: | ---: | ---: |
+| FrankenLibC | `[33.701 ns 34.193 ns 34.727 ns]` | 34.521 | 34.938 | 37.016 | 50.000 |
+| host glibc | `[26.673 ns 27.185 ns 27.776 ns]` | 27.575 | 31.903 | 34.207 | 70.000 |
+
+The focused p50 gap reproduced at `1.252x` / `6.946 ns`; mean gap was smaller
+at `1.095x` / `3.035 ns`.
+
+## Isomorphism
+
+No source code changed.
+
+- Prefix-copy length, bounded `n`, overlap behavior, destination suffix
+  preservation, byte order, and return value are unchanged by construction.
+- Ordering/tie-breaking do not apply.
+- Floating point and RNG are not involved.
+
+## Verdict
+
+NO-CODE REJECTED. Score `0.0`.
+
+The target is real enough to keep in the routing table, but every obvious
+source-level copy-panel or inline lever is already rejected. Shipping another
+manual safe-SIMD copy loop would repeat the wrong family.
+
+Next route: build or reuse an RCH-compatible assembly/codegen extraction
+artifact for safe string/memory kernels, then return only with a primitive that
+changes generated copy lowering materially: alignment-aware safe-slice dispatch,
+array-reference copy lowering with assembly proof, or an ABI-level raw-pointer
+classification lane that preserves the safe core contract.