docs: two prose copies still called the S-rung a served family

The signed-CB deletion updated the dense-gate sentence in docs/KERNELS.md
but not its copies. Neither site contains the identifier NVFP4_CB_S, so
the symbol grep that drove the deletion could not see them -- the stale
claim is written in prose ("signed S-rung").

- README.md: said a signed S-rung "is never routed through an unowned
  framework operation ... model load rejects the two unsupported FP4
  families". Wrong twice now: signed is not a family rejected at load,
  it is gone; and "the two" is one (FP4-v1).
- ROADMAP.md: the Measured-and-rejected verdict ended "the spec keeps
  them for exotic weight geometries", which docs/SPEC.md's new tombstone
  directly contradicts.

Deliberately NOT touched: docs/BENCHMARKS.md:790 ("signed-zero-only
pairs" is IEEE -0.0, not the CB family) and the dated records under
docs/audits/, which are append-only history.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TLtMbtuE5zsDKxjzZhZh8T

Author: RobTand

README.md

@@ -233,10 +233,10 @@ tile *is* an NVFP4/FP8 tile:
 Formats mix per-Linear with plain NVFP4, FP8 and BF16 inside one standard
 `safetensors` checkpoint. The 0.5 native dense serving gate is deliberately
 narrower than the format spec: CB Linears must be biasless, and FP4 must be an
-unsigned product rung with v2 scale coding. A non-`None` bias, signed S-rung,
-or FP4-v1 dense layer is never routed through an unowned framework operation:
-the public dense method rejects bias, and model load rejects the two unsupported
-FP4 families. See [`docs/MOTIVATION.md`](docs/MOTIVATION.md) for the
+unsigned product rung with v2 scale coding (the signed S-rung family was
+deleted from the runtime on 2026-08-23). A non-`None` bias or FP4-v1 dense
+layer is never routed through an unowned framework operation: the public dense
+method rejects bias, and model load rejects the unsupported FP4-v1 family. See [`docs/MOTIVATION.md`](docs/MOTIVATION.md) for the
 full argument and [`docs/KERNELS.md`](docs/KERNELS.md) for the kernel design.
 
 ---

ROADMAP.md

@@ -551,5 +551,5 @@ than silence.
 | **w2 rowpack decode schedule** | Measured negative; stays behind an environment switch as a recorded result. |
 | **Decode contract v2** (scale-epilogue hoist) | Measured **null** on the served 27B (10.10 vs 10.13 tok/s, quality-neutral) — decode is bandwidth-bound at per-byte parity, so there was nothing for the hoist to recover. Default stays v1; v2 remains available. |
 | **L2-pinned per-expert scratch pipeline** | Wedged live serving three times, including the serial variant. Removed from production dispatch and its selector surface; the underlying L2-residency hypothesis remains a historical unmeasured idea. |
-| **Signed "S-rung" formats** | Serving correctness proven bit-exact end to end, but in a matched-rate head-to-head over 776 per-(Linear, rung) comparisons the unsigned rungs won 79% of the time and the allocator placed 6 signed units against 147 unsigned. Closed as research-only; the spec keeps them for exotic weight geometries. |
+| **Signed "S-rung" formats** | Serving correctness proven bit-exact end to end, but in a matched-rate head-to-head over 776 per-(Linear, rung) comparisons the unsigned rungs won 79% of the time and the allocator placed 6 signed units against 147 unsigned. Closed as research-only on that basis; then **deleted outright** — the producer stopped emitting the family on 2026-08-17 and the runtime followed on 2026-08-23 (decode kernels, Python admission, tests, and the `NVFP4_CB_S` contract row are gone). `docs/SPEC.md` keeps a tombstone: a conforming reader MUST refuse `mode: "signed"` rather than decode it. |
 | **Retired host-branch CUDA-graph capture of the decode path** | Historical measurement, *worse*: a prefill-sized trace baked the expand arm into decode. That branch and its switch are removed. The later opaque whole-dispatch op fixed the mechanism; mode-0 `FULL_DECODE_ONLY` measured 20.1% faster on the dated close-rate 0.6B canary, pending a fresh 27B streaming gate on the current operator stack. |