fix(dsv41): enforce native required and named tool prefixes

ATTRIBUTION.md

@@ -53,3 +53,19 @@ ourselves are not listed.
 - "DeepSeek", "DeepSeek-V4.1-Flash" and "DeepSeek-V4.1-Flash-1017" belong to DeepSeek. "DGX Spark" and "GB10" belong to NVIDIA. We use the names
   only to say what this branch runs and on what hardware.
 - This branch is not affiliated with or endorsed by DeepSeek, NVIDIA, the TensorFold authors, Mia-AiLab or MiaAI-Lab.
+
+## Native DeepSeek required-tool prefix gate
+
+`zig/src/core/lanes/call_gate.zig` and its regression fixtures adapt the Python
+call gate and DSML tests from Elevatormusic/TensorFold PR #1, commit
+`003dd66273313d7d69d5675097e873e3af39ffd0`, under Apache-2.0:
+https://github.com/Elevatormusic/TensorFold/pull/1
+
+The Zig port adds request-local hooks, native lane integration, and explicit
+errors for invalid correction state. The rendered-template probe recognizes the
+separate `<` text head and `｜DSML｜` token. Existing TensorFold copyright and
+license notices remain applicable. This is prefix/name enforcement, not JSON
+argument-schema enforcement.
+
+Jayleaton's native DeepSeek grammar implementation was consulted as an
+architectural reference; no xgrammar or backend code was copied from that fork.

docs/receipts/dsv41-forced-tools/README.md

@@ -0,0 +1,112 @@
+# Native DeepSeek required and named tool calls
+
+Base: `bertholomus/TensorFold:deepseek-v41-zig-tp4` at
+`b8b115e7ed6c86d81b7575caa6f3d4ade52b323c`.
+Tested source: `253c38ca4a30fca8e0fe242d5599844e84cc187c`, Zig tree
+`12ad2fa7f1d21e85fa47129c2a5d97b8dee315a0`. The receipt amendment changes no code.
+
+DeepSeek's tool-call opener spans a text head, a DSML token and a text suffix.
+This port recognizes that split opener and enables a request-local call gate
+for native CUDA `dsv41`. It constrains the first tool-call opener and an offered
+function name across serial, speculative and concurrent generation. Named
+choices use the narrowed tool list; thinking stays unconstrained until it closes.
+`auto` and `none` do not install a gate. Upstream grammar handling is preserved.
+
+The gate checks tokens before forwarding or drafting, cuts speculative output
+at the first invalid token, and observes only committed tokens. Quoted names
+remain constrained until the closing quote. Leading whitespace before the
+opener is permitted; trailing whitespace inside its text head is rejected.
+It does not enforce argument JSON/schema conformance, later calls, or completion
+when a token limit, stop condition or cancellation ends generation.
+
+## Validation (2026-10-21 UTC)
+
+Repository-pinned Zig 0.17.0, macOS ARM64 host checks:
+
+| Command | Result |
+| --- | --- |
+| `zig test -lc zig/src/core/lanes/lanes.zig` | 132 passed |
+| `ZIG=zig bash zig/tests/server/zig_test.sh` | 375 server, five JSON and 170 engine API tests passed |
+| `zig build native -Dtarget=aarch64-linux-gnu -Dcpu=generic -j4 --summary all` | All seven steps passed |
+| `zig build test test-golden -Dcpu=apple_m1 -j4 --summary all` | 208/212 steps succeeded; 2,291 tests passed, three skipped; overall failure at inherited macOS `SO.BUSY_POLL` in `dsv41/link.zig:23` |
+| Golden corpus from that run | 309/319 equal, ten known differences, zero unexpected |
+| `python3 tools/zig/lean_check.py` | 1,615 inherited findings, unchanged from base |
+| `git diff --check` | Passed |
+
+Coverage includes required/named preflight, auto/none, unsupported-backend refusal,
+split and merged tokens, quoted-name delimiters, opener whitespace, thinking
+boundaries, request isolation, committed correction history and output limits.
+Astra reviewed the implementation and independently passed all seven focused
+gate tests. Host compilation is separate from the GPU
+checks below; the full host suite is not represented as a clean pass.
+
+## Matched GPU benchmark (2026-10-21 UTC)
+
+Four NVIDIA GB10 hosts running Ubuntu 24.04, native Zig CUDA TP4 DSpark.
+EXL3 checkpoint `bertholomus/DeepSeek-V4.1-Flash-1017-EXL3-4.0bpw` at
+`a2f91e2794923d0beebebe0d267fb3468eaf0b65`, Engram at
+`1bae0e1088993b9b82de08f59de5be5a34e42170`, patched NCCL 2.30.7.
+Both profiles use Zig 0.17.0, identical CUDA kernel assets and
+`-Dtarget=aarch64-linux-gnu.2.28 -Doptimize=fast -Dsm=121`.
+The same external build adapter exposes the existing production lane executable;
+production transport source does not differ between profiles.
+
+Effective runtime: `--context 1048576 --parallel 16 --max-tokens 32768`,
+window 2097152, round rows48, served depth solo, neighbor RDMA enabled, GID3.
+Measurements use the head-loopback API as `deepseek-v4.1-flash`.
+[Benchmark data](benchmark.json) retains source/binary hashes, per-load timings,
+OS/kernel details, request settings and the benchmark client hash.
+
+Order: upstream, patched, upstream, patched; fresh load and identical warmup
+each time. Four streams and 256 output tokens: set-a one warmup and six measured
+bursts, set-a 30-second steady admission, set-b 90-second sustained admission.
+Ordinary requests do not install a call gate. Results average two loads per build.
+
+| Workload | Upstream tok/s | Patched tok/s | Change |
+| --- | ---: | ---: | ---: |
+| Four-stream burst | 132.40 | 132.81 | +0.311% |
+| 30-second steady | 147.97 | 147.73 | -0.162% |
+| 90-second sustained | 169.84 | 169.59 | -0.150% |
+
+No meaningful performance reduction was observed in this benchmark. All 96
+measured burst replies match UTF-8 across builds; all 36 qualification replies
+match token IDs and text across builds and serial/drafted/concurrent paths.
+All workload stages report zero worker errors. Fourteen serial and four
+concurrent tool-choice feature cases pass on the patched build.
+
+The supplementary cooperative-tool workload compares auto/required/named on the
+patched build: four streams, alternating choice order, one warmup and six timed
+bursts per choice. All 84 calls return the expected function and arguments;
+output token IDs match across choices. Full-response aggregate throughput:
+
+| Choice | tok/s | Change from auto |
+| --- | ---: | ---: |
+| auto | 201.57 | +0.000% |
+| required | 200.91 | -0.327% |
+| named | 201.29 | -0.138% |
+
+Required/named throughput is within 0.33% of auto. These samples are not a
+statistical upper bound on overhead. A forced choice that changes the generated
+answer can change latency. The active-call workload is separate from the upstream
+benchmark; these checks do not establish argument-schema enforcement or
+long-context quality.
+
+Public prompts and client: [pinned `kit_bench.py`](https://github.com/bertholomus/deepseek-v4.1-tensorfold-tp4-4xgb10/blob/9af1f4567ee49ec2a31f68975982093277e7090d/tools/kit_bench.py).
+Its SHA-256 matches `workloads.client_sha256` in the data. Equivalent invocations
+(the measurement wrapper additionally retained raw responses):
+
+```sh
+python3 kit_bench.py --base http://127.0.0.1:40053 --model deepseek-v4.1-flash concurrent --set a --streams 4 --tokens 256 --reps 1
+python3 kit_bench.py --base http://127.0.0.1:40053 --model deepseek-v4.1-flash concurrent --set a --streams 4 --tokens 256 --reps 6
+python3 kit_bench.py --base http://127.0.0.1:40053 --model deepseek-v4.1-flash sustained --set a --streams 4 --tokens 256 --seconds 30
+python3 kit_bench.py --base http://127.0.0.1:40053 --model deepseek-v4.1-flash sustained --set b --streams 4 --tokens 256 --seconds 90
+```
+
+## Provenance
+
+The gate and initial fixtures were ported from an earlier native adaptation of
+[Elevatormusic/TensorFold PR #1](https://github.com/Elevatormusic/TensorFold/pull/1),
+commit `003dd66273313d7d69d5675097e873e3af39ffd0`, under Apache-2.0.
+See [ATTRIBUTION.md](../../../ATTRIBUTION.md). The scheduler integration was
+adapted to this branch's `Engine.addStream` and existing round paths. This
+change contains no transport optimization or SGLang code.