pin v2 payload: engine 2610171128001, cuda 2610160820001, cuda12 2610141455002, sbsa 2610141455003, vulkan 2610141455004, macos 2610141455005, media 2610071902001

.gitattributes

@@ -300,3 +300,5 @@ components/cuda/2610160820001/ggml-cuda-x86_64.so filter=lfs diff=lfs merge=lfs
 components/cuda/2610160820001/ggml-cuda-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text
 components/engine/2610161651001/opencoti-0.10.5-c13-2610161651001 filter=lfs diff=lfs merge=lfs -text
 components/engine/2610161651001/opencoti-0.10.5-c13-2610161651001.exe filter=lfs diff=lfs merge=lfs -text
+components/engine/2610171128001/opencoti-0.10.5-c13-2610171128001 filter=lfs diff=lfs merge=lfs -text
+components/engine/2610171128001/opencoti-0.10.5-c13-2610171128001.exe filter=lfs diff=lfs merge=lfs -text

components/cuda/2610160820001/BUILD_INFO.cuda.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610141255001; gated floor (engine-min) 2610141212001; gated with 2610160746001, 2610161700001, 2610161651001.
+Built from the engine source of build 2610141255001; gated floor (engine-min) 2610141212001; gated with 2610160746001, 2610161700001, 2610170942001, 2610171128001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|
@@ -26,6 +26,8 @@ llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a33
 llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
 ```
 
+**Gated with the r5 engine 2610170942001 (2026-10-17), the same bytes as r4:** the glibc 2.28 image on bs2 PASS; the notice gates (v2 and v3), the bug-4108 reproducer and greedy identity with r4 on the Linux library (RTX PRO 6000).
+
 **Gated with the r4 engine 2610161700001 (2026-10-16), the same bytes as r3:** the glibc 2.28 image on bs2 PASS; the bug-4108 reproducer and greedy identity with r3 on the Linux library (RTX PRO 6000).
 
 **xOllama on these bytes (2026-10-16; xOllama dev, the pin move to r3 (owner 2026-10-16 08:51), on the published bytes of index 7c1fa53f): PASS.** eleven2go RTX 3090, the Windows DLL 68cd73b1: qwen3:8b 124.5–125.1 tok/s (r2 124.2–125.1); f16-K / q8_0-V with flash attention 127.6–127.9 tok/s, prompt ~4.2k tok/s, attention on the GPU (bug-4090 fixed; the r2 DLL: 30.9); the engine gone 852 ms after a server kill. solidPC RTX 3090, the Linux library: G3 / G4 / G5 / G6 PASS (llama3 77.4 tok/s, four slots 311.2, council 6 / 6, speech 8 / 8, video 33 frames).

components/engine/2610171128001/BUILD_INFO.engine.md

@@ -0,0 +1,61 @@
+# engine 2610171128001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `opencoti-0.10.5-c13-2610171128001` | any | bin | `98836c63f84b59dfa5abc90fef5b637470a75a0fe3fe622de12b32af1ec11540` | 145676430 |
+| `opencoti-0.10.5-c13-2610171128001.exe` | win-x86_64 | bin | `98836c63f84b59dfa5abc90fef5b637470a75a0fe3fe622de12b32af1ec11540` | 145676430 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+ABI `media` = `b099409b3029a6e380268e31d1caefa6706569961d32e72b3543fbc73903d95c` (exported) — sha256 of this list:
+
+```
+llamafile/oc-audiocpp/audiocpp.h 6ad0b55d5f28adc9ad60a45bf9a7caa428c4f5441389a6932d6670a8e2a90d73
+llamafile/oc-codec/oc_codec_sidecar.h f2c9001d1414aa2a55533b867b79c52a52c682558e0234c5bebd30bf97b91005
+```
+
+ABI `loader` = `d6ff34022f1d3fb5af56ca0c80e3f4e552bf09f1bebdf0cb5f314a454bf38dbc` (exported) — sha256 of this list:
+
+```
+.cosmocc/4.0.2/bin/ape-m1.c 78af79e20abbb3f99a97355959b59bd58949564064adb754f0302ebabb4bf7a3
+```
+
+**c13 r5 dev snapshot — the development build of the c13 r5 release, for integration testing.** Engine 2610171128001 (`opencoti-0.10.5-c13`, dev channel) is built from the same source as the c13 r5 release engine 2610170942001 (chain 0001–0693; identical by content). The libraries are the r5 release index's — the r4 bytes, unchanged. This engine itself ran the glibc 2.28 route gate (Vulkan + CUDA), the notices v2 gates and the notices v3 gate on bs2, the copy-bar switch check included; the scoped regression gates of the release ran on 2610170942001 (see the release notes). The r4 dev snapshot was 2610161651001.
+
+What is new in r5 (chain 0690–0693):
+
+- **Generation notices contract v3** (0692, c13 row 46.1), opt-in per request with `"notices": {"version": 3, …}`; v2 stays the default and is unchanged. A v3 notice is one plain-text line in the model's own voice, planted as its own paragraph at a prose boundary: no `<engine_status>` frame and no bars. The built-in thinking lines carry the words left ("I'm stopping my thinking now and writing the answer — only {words} words left."), the answer lines do not ("I'm writing the conclusion now."). Every planted line is reported exactly, so a client can hide it: a top-level `notice` field on the stream chunk that carries the line's first token (`index`, `phase`, `channel`, `text`, `break`, and `cut` when the line was cut short), the list in `timings.notices`, and the built-in forms in `/props` `generation_notice.v3`. Routes: `/v1/chat/completions`, `/v1/completions`, `/completion`. Feature `generation_notice_v3`. Measured with the built-in v3 lines on the development build of this source (bs2, four models: Qwen3.6-27B, Nemotron 3.5 Lightning, Qwen3.5-9B, Gemma 4 E4B; answer cap 3000 with prompts asking ~3000 words, 24 answers each): answers that end inside the cap 13 of 96 without a notice, 64 of 96 with the v3 answer lines (per model 1 → 13, 1 → 20, 0 → 8, 11 → 23); no copy of a planted line in any answer.
+- **The copy bar** (0693, c13 row 46.3, bug-4115), v3 only. A model can copy a line the engine planted: Qwen3.6-27B wrote a planted line again, verbatim, in about 1–2 % of thinking requests (3 of 168 with MTP self-speculation, 1 of 72 without), once or in a loop of up to 21 copies. A copy has no record, so a client hiding by the records shows it. Once the line being written would become the first 8 or more characters of a line planted earlier in the response, the token that continues the copy is refused, and the engine log names it. Measured on the development build: 0 copies in 216 Qwen3.6-27B-MTP thinking requests; thinking cut at the budget 21 against 22 times on the same 120 prompts and seeds; finished answers 64 of 96 with and without the bar; the same decode speed. On by default in both phases; `notices.copy_bar` switches it per request: `true`, `false`, or `{"think": bool, "answer": bool}` (a missing phase stays on; on a v2 request it is a 400). `/props` `generation_notice.v3.copy_bar` shows the defaults; feature `generation_notice_copy_bar`.
+- **The notice bar no longer writes random tokens** (0690, bug-4112 / bug-4113, in every c13 build since r2). On the device sampling path a token refused by the bar was resampled from an all-zero logits row, i.e. at random from the whole vocabulary, which could write stub lines such as `<engine8_status …>` and start a loop of them. Now the engine keeps the row for every request with notices, and the built-in v2 wording is the model's own voice (S2), still in `<engine_status>…</engine_status>`.
+- An experiment-only environment knob, `OPENCOTI_NOTICE_X_PLAIN` (0691), off by default.
+
+What is new in r4 (chain 0689):
+
+- **No placeholder token in the output where the notice bar acts** (0689, c13 row 45.1, bug-4108). Gemma 4's 6227 `<unusedNN>` tokens are ordinary tokens in its GGUF files; with notices on, Gemma 4 26B-A4B's own copy of a notice reached the reply as `<engine_<unused29>_status output_budget="140">…` (the built-in form) or `<system_notice` with `<unused22>` inside (a request template), or as a stray `<unused30>` right after a planted notice — visible, and not hidden by a client that hides from the notice's opening. Now a placeholder (marked unused, or named `<unusedN>` / `[unusedN]`) is never sampled while the bar is on, a line may not start with any template's opening character at a prose boundary after a notice, and the line right after a planted notice gets the same bar. Measured on the release bytes (2610161700001) (bs2, Gemma 4 26B-A4B Q4_K_M, the built-in form and a self-closing template, 48 answers each): 0 placeholders and 0 copies (r3: 2 + 1 and 9 + 12), every planted notice in the text.
+
+What is new in r3 (chain 0666–0688):
+
+- **The CPU at stock llama.cpp speed** (0666–0675, 0678, 0685–0687, c13 row 41.1). Measured on the release bytes (2610160746001) against stock llama.cpp b11521, Qwen3.6-35B-A3B UD-IQ3_XXS on the CPU: EPYC 9575F 6 threads **110.2 %** (r2 92.3 %), 1 thread 107.4 % (r2 88.3 %); Mac mini (Apple M-series, Metal off) 6 threads 99.7 % (r2 81.7 %), 4 threads 97.9 % (r2 77.4 %); Ryzen 5 5600G 6 threads 102.2 % (r2 96.1 %, a noisy host); Raspberry Pi 5 (Qwen2.5-1.5B, 4 threads) 104.2 % (r2 97.3 %). The causes, each fixed with stock's own code path: cosmopolitan's `memcpy` copied backward (a forward copy replaces it); flash attention runs on the CPU as in stock (llamafile's layer turned it off); stock's tiled int8 matmul for K / IQ quant prompts; stock's Q8_0 / Q4_0 prompt kernel (with an AVX-VNNI copy); f16 V rows converted at the variant's ISA; the CPU work buffer 64-byte aligned; F16 / BF16 prompt matmuls reach stock's kernel; ARM kernels built with clang and tuned for Apple cores; IQ2_S matrix × vector on ARM through stock's kernel; no OpenMP (ggml's own threadpool, as stock's portable builds); a CPU-only engine samples on the host; on macOS a recurrent state on the CPU is f32. `--recurrent-state-type {bf16,f16,f32}` (0680).
+- **CPUSet core selection** (0671, row 41.2; a port of the owner's llama.cpp PR #6832): threads are pinned, the default pick is physical cores, no efficiency cores, one last level cache (the prompt pool gets every last level cache); Windows CPU sets across processor groups (no 64-core limit). `--cpuset on|off`, `-bco`, `-llct`, `-atc`, `--cpuset-skip-core0`. Thread placement had never run in an opencoti build (bug-4080). i5-14600K: 115.2 % of stock native, 9950X3D 103.7 %.
+- **cpp-httplib 0.60.1** (0676, 0688, row 41.3): with stock's build definitions — a request body over 8 KB without a JSON content type is accepted (HTTP 413 before, bug-4082); a realtime websocket message that arrives in pieces no longer ends the session (bug-4105).
+- **Generation notices, contract v2** (0681, 0682, 0684, row 41.4; feature `generation_notice_v2`): the request's `notices` object (steps by `at_percent` or `left`, templates with `{left}` / `{total}`), `--notice-templates FILE`, `--print-notice-config`, `/props` `generation_notice` (the templates and the spans a client hides), `timings.notices_planted`, `POST /v1/notice {key, kind: "message_waiting"}` (through the router by key). A notice is planted only at a prose boundary, never inside code, and only while 200 or more tokens of its phase are left.
+- **`num_ctx_hold`** (0683, row 41.6; feature `elastic_window_unheld_v1`): `false` sizes the window from `num_ctx` / `num_ctx_min` and releases it at the end of the request, with a `session_id` too.
+- **Tool calls keep Windows path backslashes** (0679, bug-4092).
+- **The Windows CUDA library runs flash attention with mixed K / V types on the GPU** (bug-4090): every Windows DLL since c10 r3 was built without `GGML_CUDA_FA_ALL_QUANTS`, so an f16-K / q8_0-V cache ran attention on the CPU. New cuda component 2610160820001.

components/engine/2610171128001/opencoti-0.10.5-c13-2610171128001

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98836c63f84b59dfa5abc90fef5b637470a75a0fe3fe622de12b32af1ec11540
+size 145676430

components/engine/2610171128001/opencoti-0.10.5-c13-2610171128001.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98836c63f84b59dfa5abc90fef5b637470a75a0fe3fe622de12b32af1ec11540
+size 145676430

components/vulkan/2610141455004/BUILD_INFO.vulkan.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610141255001; gated floor (engine-min) 2610141212001; gated with 2610141255001, 2610141633001, 2610160746001, 2610161700001, 2610161651001.
+Built from the engine source of build 2610141255001; gated floor (engine-min) 2610141212001; gated with 2610141255001, 2610141633001, 2610160746001, 2610161700001, 2610170942001, 2610171128001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|
@@ -26,6 +26,8 @@ llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a33
 llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
 ```
 
+**Gated with the r5 engine 2610170942001 (2026-10-17), the same bytes as r4:** the glibc 2.28 image on bs2 PASS; the Linux library.
+
 **Gated with the r4 engine 2610161700001 (2026-10-16), the same bytes as r3:** the glibc 2.28 image on bs2 PASS; the Linux library.
 
 **xOllama on these bytes with the r3 engine 2610160746001 (2026-10-16; xOllama dev, the pin move to r3 (owner 2026-10-16 08:51), on the published bytes of index 7c1fa53f): PASS.** eleven2go (Windows 11): RX 9070 XT, pinned by name, 100.5–103.3 tok/s, 103.9 after 120 s loaded idle, the card OK, 0 LiveKernelReports; Radeon iGPU 5.0 tok/s, 37 / 37 layers. The Linux library: not run by xOllama for r3.