fix: harden miner recovery, pool validation, and launchers

[SWvheerden]

Summary

This PR hardens miner recovery, pool handling, and the multi-GPU launchers:

• Stops surviving GPU workers when one worker fails and propagates that worker's exit code. The Linux launcher reports a missing backend as exit 5 and escalates TERM to KILL after a bounded grace period.
• Separates explicit login rejection, clean disconnect, connected no-job silence, and malformed protocol data.
• Exits with code 6 after repeated connected no-job timeouts and code 7 after three consecutive protocol failures. A protocol streak resets after a valid job update or a clean connection, including the immediate update-then-error race.
• Validates complete JSON documents, including container matching, escapes, numbers, commas, and trailing data, while retaining status-object share acceptance.
• Bounds outstanding submit tracking and removes failed CUDA recovery slots so zeroed proofs cannot reach host verification.
• Rejects empty, whitespace/control, and overlong wallet/login values. Ambiguous Base58-length pool usernames now warn without being blocked, and the bounds cover Base58, hex, emoji, and pipe-grouped Tari encodings.
• Documents supervisor-facing exit codes 0-7 and adds Linux and Windows launcher regressions to CI.

Validation

• Windows MSVC and Ubuntu 22.04 g++ protocol/reliability suites: pass.
• Windows PowerShell and Linux shell launcher regressions: pass, including failure propagation and cleanup of a TERM-resistant survivor.
• CUDA 13.2 Windows build matrix: all nine release solver, pool miner, and reference targets pass for sm_86, sm_89, and sm_120.
• RTX 5090 exact recall: five 4,200-graph runs (21,000 total) pass candidate repeatability, reference repeatability, candidate-vs-reference, and pipeline-4 parity. Every run found the same 110 proofs with zero verification failures.
• Fresh final sm_120 smoke: 100 graphs, two verified cycles, zero verification failures.
• GitHub Actions for commit 9b987d8: all three jobs passed.

Deliberate boundary

The recall matrix proves exact solution correctness; it is not a throughput comparison. No graphs/s A/B against v1.1.2 was performed.