stratum: harden and streamline BLAKE2b share submissions

[f4u57x]

Summary

• cache the hidden previous-block commitment once per BLAKE2b job and reuse it for submitted shares
• strictly validate hexadecimal job IDs, extranonce2 values, times, and nonces before interpreting them
• add unit and Stratum regression coverage for cached work-header construction and malformed submissions

Rationale

The hidden previous-block value is constant for every share on a job. Computing it during job refresh removes a tagged SHA-256 operation from the share-submission hot path, at a storage cost of 32 bytes per job (8 KiB for 256 jobs).

The existing lookup-based conversion maps invalid hexadecimal pairs to zero. Rejecting malformed fields explicitly prevents distinct invalid submissions from aliasing valid zero bytes and avoids unnecessary proof-of-work processing.

Payout and address-validation behavior is unchanged.

Testing

• normal CMake build
• full datum_gateway --test suite
• AddressSanitizer and UndefinedBehaviorSanitizer build and full test suite
• git diff --check

[jasonsopko]

Tested ACK fc5ec5c

Ubuntu 24.04, gcc 13.3, clang 18.1, on top of master b9ea7dc:

• The CI matrix run locally: gcc and clang with -Wall -Werror, ENABLE_API=OFF, DATUM_API_FOR_UMBREL, ASan+UBSan. All build clean and pass --test. git diff --check clean.
• Negative control: the new datum_blake2b_malformed_submit_hex_tests applied to master without the code change fails five of its six cases, so the test does catch the bug. Master decodes the bad extranonce2, ntime and nonce to zero bytes, hashes them, and answers H-not-zero; only the bad job id was caught, by the index check. With the code change all six answer unknown-work before any hashing, as described.
• Both writers of a BLAKE2b job's prevhash (update_stratum_job and the coinbaser path) go through datum_stratum_job_refresh_blake2b, so blake2b_prevblock_hidden is set before a job is published, and the submit path reads it only in the BLAKE2b branch. Every new decode follows the existing length checks, so nothing reads past the string. Dropping strtoul also drops its acceptance of a leading 0x, whitespace or a sign, which no miner sends.

nit: datum_pow_decode_u32_hex_exact could be datum_pow_decode_hex_exact(hex, 4, buf) plus a big-endian unpack, keeping one decoder. Not blocking.

Not run with live miners.

[luke-jr]

This PR should really have at least 4 or 5 independent commits

[luke-jr]

Belongs in utils

[f4u57x]

Moved into datum_utils as the generic hex_to_bin_exact helper. The PoW-specific decoder has been removed.

[luke-jr]

Probably should take inspiration from hex_to_bin

[f4u57x]

Updated. hex_to_u32 now reuses the shared exact byte decoder and performs only the big-endian unpack. Tests cover uppercase, lowercase, mixed-case, truncated, overlong, and non-hex input.

[f4u57x]

Updated in response to review:

• Split the change into five independently reviewable commits.
• Moved strict hexadecimal decoding into datum_utils.
• Based 32-bit decoding on the shared exact byte decoder.
• Separated malformed job/extranonce validation from time/nonce validation.
• Kept BLAKE2b prevblock caching in its own commit.

Verification: fresh normal and ASan/UBSan builds and full test suites pass, and git diff --check is clean. Payout and address-validation behavior remains unchanged.

@luke-jr, would you please take another look?

[jasonsopko]

Tested ACK a35e10f

Re-review after the split. The six commits add up to the same change I ACKed at fc5ec5c plus what Luke asked for and one fix: the decoder now lives in datum_utils as hex_to_bin_exact, hex_to_u32 is a big-endian unpack on top of it (the nit from my first pass), and hex_to_bin_exact checks the first nibble before reading the second so a short string is rejected inside its allocation. Nothing else moved.

Ubuntu 24.04, gcc 13.3, clang 18.1, on master b9ea7dc:

• Each of the six commits builds with -Wall -Werror and passes --test on its own.
• The CI matrix on the head (gcc and clang -Werror, DATUM_API_FOR_UMBREL, ENABLE_API=OFF, gcc ASan+UBSan) builds clean and passes --test. git diff --check clean. A clang UBSan-only build fails --test at datum_protocol_tests.c:133, the misaligned store that is on master and that #14 fixes; not this PR's.
• Negative control: this head with datum_stratum.c reverted to master fails five of the six malformed-submit cases and the cached-prevblock check in the refresh test, so the tests catch both changes.
• Over a real socket on regtest (BLAKE2b from height 1), twelve mining.submit lines against a gateway built from this head and one built from the same tree with master's datum_stratum.c. The master path answers H-not-zero to all twelve: a job id whose coinbase byte is g0, an extranonce2 ending in g, 0x000000 and 0000000 as ntime, -0000001 as nonce, all decode to something and get hashed. This head answers unknown-work to those nine before hashing and H-not-zero to the three well-formed ones, uppercase hex included.
• fuzz_stratum_line, a libFuzzer harness of mine that feeds one JSON line to the Stratum command handler for an authorized miner with a job present, ASan+UBSan, on this head for 10 minutes, 73.6M inputs from the seed corpus and dictionary, no findings.

#15 carries this change in innerhat's restructured form. Whichever goes first is Luke's call: if this one does, #15 rebases onto identical hunks; if #15 does, this closes.

Did not test with live miners.