[helly25]

Compile static hash-field qualifiers once, reducing repeated parsing when rendering templates. Dynamic defaults, copied templates, snapshot lifetimes and output bytes retain their existing behavior.

AG;DR

Dependencies #1068 and #1069 are merged. The base is current main (6d6a2c9abf972ffffd064939381a1df2f40c2776). Fresh main-based CI on d95860e8 found insufficient changed-line coverage; the refinement below removes the obsolete parser and adds direct classification/default-usage tests. No threshold is weakened and no executable lines are excluded. Fresh current-head CI, not earlier PR contexts, determines readiness.

Implementation:

• A private two-byte HashBinding stores a shared-catalog algorithm ID and explicit/default encoding choice beside the existing segment post-processing enum. No extra segment string, borrowed qualifier view, allocation or second algorithm catalog. Tracked template storage is identical for every measured shape; this is conservative capacity accounting, not an RSS claim.
• Explicit algorithm/encoding choices avoid repeated qualifier parsing. Context-dependent defaults resolve each render, including invalid defaults and changes after compilation; independent explicit choices ignore unrelated defaults. Copies/moves, transformed fields, extraction, static diagnostics, derived-content requests and cached-digest ownership/invalidation remain covered.
• Bind bare/transformed defaults directly and validate bound native qualifiers once. Use the existing nonzero binding tag for hash dispatch; ordinary fields retain their direct renderer calls. Supplied snapshots lend digest views; the local fallback returns owned data.
• Three initial tests cover changing defaults across copies/moves, every shared-catalog algorithm after temporary input storage expires, and invalid dynamic defaults with independent explicit fields/transforms. Existing field, extraction, derivation, lifetime, snapshot, output and fuzz tests remain.
• Add common explicit/partial/default warm/cold field and compilation benchmark cases. No public flag/API, dependency, worker, lock, new content cache or host-I/O path.

Measurements and costs:

• Both hosts retain a first two-variant comparison (47 cases x three sessions x two participants), then a refined three-variant comparison (47 x three x three). Preserve populations separately. Reused frozen reference/first binaries have verified exact hashes and identical compiler/configuration. Every memory case keeps fastest seven of nine with independent fixtures/warmups; batching amortizes only untimed metadata startup.
• Independent native confirmation covers 20 cases x three sessions with baseline/selected/byte-identical baseline repeat: 1,800 full-output/exit-checked invocations per host including discarded warmups. It covers 10/1,000/10,000 files, long paths, listing/hash fields, and one/three requested workers. Separate memory confirmation covers 16 cases x three sessions x three participants. No same-host timing overlaps builds, lint or instrumentation.
• Explicit cached hash kernels improve about 41-56% in confirmation. Selected larger whole-engine in-memory commands improve about 2-5%; native commands improve much less, generally fractions of a percent on macOS and up to roughly 3% in selected Linux cells. This is not a universal rendering-speedup claim.
• Linux simple-name field rendering retains about a 2% cost, roughly one nanosecond per entry; explicit-hash compilation is about 4.4% slower in confirmation. macOS cold/non-hash controls vary, including a 6.8% slower name-template compilation cell. All raw initial/confirmation controls, absolute times and reference-repeat movement remain visible.
• The first prototype duplicated qualifier validation and separately loaded a dispatch tag. Its initial hash-compilation regressions of roughly 29-49% are rejected. Preserve its complete source, data and setup/lint failures separately; no complexity threshold or policy was weakened.
• The first qualified full unstripped CLI footprint increases 1,440 bytes on Linux and 368 bytes on macOS; the final direct-tag refinement is measured separately below.

Verification:

• All 313 selected root-and-extension release suites pass on each host; Linux requires working FUSE. Nine focused ASan suites pass per host; nine MSan and nine TSan suites pass on Linux.
• Actual scoped clang-tidy passes the changed implementation, test and common benchmark on both hosts; the refinement passes a focused implementation recheck. Exact four-file guards pass after source-swapping builds and before/after complete verification. Normal commit/push hooks also check affected header consumers. Separate three-TU checks per host explicitly include the changed XFF header and pass; the existing inherited MBO-only header-filter defect remains tracked independently in TODO.md.
• CHANGELOG, TODO, performance analysis and release-site mapping link docs/benchmarks/hash-field-binding-20261103.md. Its 1,148,942-byte lossless archive has SHA-256 e5e1a4c0468510bcb2daaf2dda872405513435e51c7475f60c3e594e7aa5db35; it retains raw timings, source/compiler/configuration/link/binary identities, scripts and validation logs.
• Current-head GitHub checks determine merge readiness. O22c2 reusable output/streaming and all other retained performance candidates remain open.

Final coverage-driven refinement:

• Delete the obsolete generic hash parser, which was never called after binding and served only as a function-identity tag. The existing binding ID distinguishes hash fields; ordinary renderer calls and segment storage remain unchanged. Centralize field-versus-literal classification for rendering and extraction. No duplicate catalog, extra object field or artificial coverage exclusion.
• Two additional regressions distinguish literal/escaped/unknown/hash segments and exercise explicit/default/partial/transformed/invalid default usage. Existing all-catalog algorithms, copied/moved templates, extraction and ownership tests remain. The final test-only addition passes release/ASan on macOS and release/ASan/MSan/TSan on Linux.
• All 313 release suites pass per host; nine focused ASan suites pass per host and nine MSan/nine TSan suites pass on Linux. Three explicit changed-header consumers pass per host, followed by a focused recheck of the final regression. Focused Linux changed production field coverage is 86/88 lines (97.73%) and 67/70 branches (95.71%). The full CI coverage gate remains required.
• Measure old/tag variants separately: 16 memory and 20 native cases x three sessions x two participants on each host, fastest seven of nine with discarded warmups and complete native output/exit oracles. Sixteen warm explicit hash fields improve another 7-8%; larger whole-engine repeated fields improve 0.9-3.2%. Costs remain visible: a Linux single warm explicit hash cell is 13.1% slower, macOS ordinary-name/component kernels are 5.1%/3.3% slower, and compilation/native controls vary. Native changes range -2.0% to +3.5% Linux, -0.6% to +0.8% macOS. Do not pool these populations with the original baseline comparison.
• Full unstripped CLI decreases 4,896 bytes Linux and 208 bytes macOS relative to the preceding bound-field implementation. No same-host release timing overlaps builds, tests, lint or instrumentation.
• docs/benchmarks/hash-field-binding-tag-20261103.md records the refinement. Its 585,438-byte lossless archive has SHA-256 558057de3b7c3f6cb8f84632721ad0297deae2e8bd050c7edd2099ba2ef9ebd6; it includes sources, manifests, scripts, raw timings, validation and coverage logs. Preserve the first artifact independently.