[helly25]

Hash primaries now decode explicit algorithm and encoding syntax once, reducing repeated work in CPU-bound hashing and verification. Expected templates, live defaults and conditional reads preserve their existing behavior.

AG;DR

Based on main after #1055 merged. Synchronization head 4e77548bebf05152bb55b64dadd8e7ec9da8f539 contains main 00d4f84df01e00f1ce2d7c83576ac38b2eb2fb7b; the complete implementation tree is unchanged from measured head 327ea789f840a42712cb92ddf29db52bd812e7e3. Normal synchronization commit/push hooks pass. Required CI is running on the synchronized head. The earlier coverage job failed during LLVM extraction with runner disk exhaustion, before tests; the new synchronized run replaces that failed context.

Implementation:

• Add a compact prepared hash record using an index into the existing AlgorithmNames registry, encoding/default-use/validity bits, and the existing unique expected-template owner. No second algorithm list or public hash API is introduced.
• Assert that the record fits the existing 24-byte operand alternative; unrelated operand storage does not grow. Bare hash acquires a prepared pool record/capacity, so preparation is not allocation-free.
• Register expression-aware preparation in the existing constexpr dispatch. Argument-only families use one preparation adapter and preserve their previous missing-argument guard. A boolean uses existing dispatch-record padding rather than adding another callback pointer.
• Resolve only requested defaults from the reached direct context; no borrowed source/default view is cached across calls. Expected-field rendering still precedes conditional hash resolution/reads, and verification bookkeeping remains unchanged.
• Preserve direct-tree reference callbacks. Add no per-file lock, worker, clock, shared ownership, flag, mutation route or exception mechanism. Hash/verification coordinator admission is unchanged.

Evidence and tradeoffs:

• docs/benchmarks/prepared-hash-operands-20261102.md and compressed lossless data retain source/binary/fixture identities, compiler versions and hashes, build configuration, actual CLI link actions/parameter contents, scripts, logs, private pilot and final/confirmation samples.
• Per host: 56 native cases x three sessions; 36 in-memory whole-engine cases, 16 hot kernels and eight preparation-only cases x three sessions; confirmation with an identical baseline adds eight native cases x five sessions and 12 memory cases x three sessions. Discarded warmups and fastest seven of nine remain separate contracts.
• Explicit hashing/equality improves roughly 6-9% in in-memory confirmation on both hosts. Native commands stay near baseline because file I/O dominates. Defaults improve less; no universal speedup is claimed.
• Preserve cold costs: a one-primary hash preparation adds roughly 50-70 ns, and Linux name preparation adds about 27 ns. The report shows absolute values, percentages, duplicate-baseline controls and unrelated name/diff results; their improvements are not credited to hash parsing.
• Unstripped CLI: Linux +27,176 bytes; macOS +7,584 bytes. Existing digest/output copies and field-hash rendering remain O22 work; wider EP03/EP07 qualification is still open.

Validation:

• Complete extension-enabled release suites pass 322/322 on macOS and Linux, with required Linux FUSE mounts.
• Six focused ASan suites pass on each host; six Linux TSan suites pass.
• Dedicated tests cover every registered algorithm and both encodings, reused/moved plans with changing defaults, invalid/empty direct defaults, malformed specs with content-dependent expected fields, missing arguments, empty/binary input, failed reads and skipped actions.
• All four modified translation units pass scoped local clang-tidy after addressing const/optional/test-helper findings. Normal enforcing commit/push hooks pass before PR creation; readiness still requires current pushed-head CI.

Reconcile O15b into the canonical performance analysis, TODO, expression design/qualification audit, changelog and release-site mapping. Keep field-hash binding and broader template-family qualification explicitly open.