Keryx Miner Supr v0.16.7 — large private-prompt compatibility

• Port upstream keryx-miner v0.6.2-PoM's traced context-overflow retry to both
  CUDA and Vulkan. Try the full decoded prompt first; only a context overflow
  can trigger one UTF-8-safe head/tail retry within the model's budget. Keep
  the original prompt length, request prefix, serving card and audit trace.
• Match upstream's optional trace.prompt_len default and consensus clamping:
  count the original opened bytes before decompression, never inflated text.
• Preserve upstream's 1 MiB DEFLATE byte cap before lossy UTF-8 conversion;
  admit its bounded conversion expansion on private traced requests.
• Reply explicitly to unservable pool dispatches using upstream error 46
  (malformed requests: 20) when an id is supplied, so pools can redispatch.

Pool operators must also remove obsolete 64 KiB decoded-text routing limits
and forward the original opened bytes. v0.16.4 already introduced compressed
pool-prompt decoding; v0.16.7 adds the retry and admission/refusal fixes.
See the integration contract.

H16 consensus, model weights, native engines, GPU mining kernels and per-card
pause/proof requirements are unchanged. No new hashrate gain is claimed.
Existing physical RDNA1, Windows GPU and ten-card qualification gaps remain;
macOS retains its experimental Metal/H15/H16 limitations.

Packages: Linux NVIDIA modern and legacy, Linux AMD, HiveOS, mmpOS, SMOS,
Windows NVIDIA and AMD, macOS arm64, and Docker
ocminersupr/keryx-miner-supr:0.16.7 (latest). Licensing remains
Source-Sharing License 1.0; upstream licenses and previous grants are preserved.

Validation: 285 host/integration tests, 90,647 byte-exact comparisons with upstream's
UTF-8 head/tail rule, and 33 Linux package startup checks on Ubuntu 20.04/22.04/24.04.
Final Linux packages passed isolated H16 traced inference, all audit opening phases,
custody upload/restart/replay and mining resumption on RTX 5090 and RX 7900 XTX.
Synthetic probes exercised 74,739/79,150 decoded bytes, tier-4 Kimi at 500 KiB, and a
compressed input expanding beyond the 1 MiB cap (pre-AVX Vulkan + descriptor path).
These tests forced an 8,192-token context to exercise the retry; they are not a
replay of the original user's private prompt contents. Error 20/46 replies and a
valid unservable-context refusal were also exercised. Live TLS pool checks accepted
61 NVIDIA and 5 AMD shares with zero rejects. These are functional checks, not
new performance measurements or on-chain AI request inclusion claims.

The independent audit verifier's 22 Rust source files match official node
v1.6.7-PoM byte-for-byte. Physical RDNA1/RX 5700, Windows GPU execution and ten-card
hardware remain unqualified. Existing benchmark tables retain their original
version/workload labels.