Harden wallet errors, solver watchdogs, and Tari consensus tests

[JustAResearcher]

Closes #9
Closes #12
Closes #13

What changed

• Reject wallet values containing whitespace/control bytes before connecting, while preserving Tari's multiple valid address/login formats.
• Track unique submit request IDs so login errors cannot inflate rejected-share telemetry, tolerate valid JSON whitespace, retain nonterminal submits, sanitize pool diagnostics, and exit 4 after three explicit login rejections.
• Propagate CUDA setup, trim, copy, synchronization, and recovery failures through both standalone and pool-miner paths.
• Add a per-context watchdog that exits 5 after three consecutive zero-edge trims, including pending work drained during pool disconnect/reconnect.
• Fix the pipelined same-context recovery race by completing recovery and consuming solutions before relaunching that context.
• Pin Tari key derivation, edge construction, proof packing/hash, and structural verifier behavior with 22 offline assertions.
• Enforce LF for HiveOS .conf files so the shell wrapper can source the manifest and preserve miner exit codes.
• Add Windows/Linux reliability tests and a CUDA 12.8 CI compile matrix for every shipped solver/pool architecture (sm_86, sm_89, sm_120).

Why

Invalid wallet replies were treated as network timeouts and share rejects. Runtime CUDA failures could collapse into an endless zero-yield loop that stayed connected and exited 0. The self-test also checked round trips/determinism without pinning Tari consensus outputs. During review, the release prelaunch path was additionally found to overwrite recovery state before a rare cycle was recovered.

Validation

• Windows MSVC and Linux g++ protocol/reliability suites: pass.
• Tari wrapper self-test: 22/22 on Windows and Linux.
• Windows CUDA 13.2: solver and pool miner build for sm_86, sm_89, and sm_120.
• Linux CUDA 12.8 / Ubuntu 18 compatibility build: all six binaries build, launch, contain the intended cubin, and remain compatible with GLIBC 2.14.
• RTX 5090: deterministic 128-graph pipeline 1/4 runs produce the same four verified cycles with zero verification failures; compact and whitespace-formatted loopback pools accept shares; forced zero-yield and a real CUDA API error exit 5 without orphaned miners; repeated spaced result: false login replies exit 4 without no job received noise.
• mini17 / RTX 4070 Ti SUPER Linux: healthy 128-graph pipeline 2 run reproduces the same four cycles with zero verification failures; forced zero-yield and real CUDA API error exit 5; GPU returns idle and Vast services remain running.
• No verified idle sm_86 hardware was available, so that profile is compile-gated but not live-run in this PR.

The CUDA CI container is a compile gate only. Release assets must continue to come from the Ubuntu 18 compatibility build path.