Neptune mining session contract (v0.15.1)

Consensus is selected from the exact pinned network/height schedule, never by whether a PoW field happens to decode. Beta and Gamma use fast PoW; pre-Beta rules use their corresponding leaf prefix and ordering. Required version and lustration fields are checked before work. RPC failures and malformed templates must not trigger allocation of a legacy mining buffer.

A session has one attempt budget across CPU/GPU workers, template refreshes and submission retries. Reservations are bounded, non-overlapping, overflow-safe. An exhausted device is not an unavailable device. Explicit CPU/GPU selections are honored; auto may select CPU if GPU initialization is unavailable.

The host owns the accepted result under one mutex. GPU mapped state is private to an exclusive, synchronous dispatcher. CPU workers never access GPU result flags or payload. Only a completed, CPU-validated GPU result may enter the host winner slot, which is write-once. No borrowed mapped data escapes the dispatcher.

Mining proceeds in bounded batches and refreshes the template between batches and before submission. A changed parent invalidates a pending result. Submission rejection refreshes the template and continues within the original budget. No template is a bounded error, not an unbounded busy wait. Network requests have a timeout. Legacy buffer construction remains explicit, expensive work; bounded unit/oracle tests must not allocate the production-size buffer.

Validation includes deterministic reservation/publication/refresh tests and small real Metal/native hash and threshold comparisons. No mock acceptance is a claim of live mining or successful network submission.

Implemented CLI behavior

--backend cpu and honeycrisp select CPU mining (the latter is an explicit alias for the host CPU implementation). gpu selects Metal only and reports unavailability; auto uses a reusable combined CPU/Metal miner when available, otherwise CPU. There is no second full-budget retry after exhaustion.

The CLI checks node_network on the same HTTP endpoint before each template fetch. Mainnet boundaries are 15000 / 23401 / 38000 / 40300; testnet-0 boundaries are 120 / 3571 / 3669 / 4650 (Reboot, Alpha, TVM1, Beta, Gamma). Regtest, testnet-mock and other valid testnet IDs use Gamma, matching upstream.

Templates require matching header/metadata parent and version. Fast PoW inserts the declared header version in pathA[26][4] and validates all six lustration codec fields against exact i128 JSON metadata (no floating-point conversion). The parent remains fixed throughout a batch of at most 4096 reserved attempts. A fresh template is fetched before submission. Rejected/stale results consume their reservation; retries never reset the global attempt cap or nonce offset. HTTP requests time out after 15 seconds. A null template or transport error is reported explicitly. This is a bounded batch refresh, not asynchronous live tip streaming; a tip can still change between the final check and submission, which is handled as an ordinary rejection.

Legacy Reboot builds leaves from the current MAST commitment without bit reversal. Alpha/TVM1 use the parent digest and bit-reversed order. Buffers are cached only while that exact prefix/order matches, and old buffers are dropped before rebuilding. Construction itself is not asynchronously cancellable and requires approximately 43 GB peak memory; full-size construction remains an operational gate on suitable hardware. No release test pretends a reduced tree is a full 2^29-leaf live mining run.