helly25 · Nov 3, 2026 · Nov 2, 2026 · Nov 2, 2026 · Nov 2, 2026 · Nov 2, 2026
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -24,7 +24,7 @@ repos:
     hooks:
       - id: check-added-large-files
         # Retain the exact lossless cross-platform performance evidence archives.
-        exclude: ^docs/benchmarks/((executor-copy-audit|bounded-read-ahead)-20261101|(directory-task-slots|packed-visit-snapshots|bounded-match-results|useful-match-workers|direct-evaluator-workers|derived-content-handoff|immutable-default-catalogs|prepared-diff-configuration|prepared-hash-operands|aggregate-comparison-totals|byte-shingle-representations|archive-input-buffers)-20261102|(digest-and-record-storage|caller-owned-streams|direct-source-materialization)-20261103)\.json\.gz$
+        exclude: ^docs/benchmarks/((executor-copy-audit|bounded-read-ahead)-20261101|(directory-task-slots|packed-visit-snapshots|bounded-match-results|useful-match-workers|direct-evaluator-workers|derived-content-handoff|immutable-default-catalogs|prepared-diff-configuration|prepared-hash-operands|aggregate-comparison-totals|byte-shingle-representations|archive-input-buffers)-20261102|(digest-and-record-storage|caller-owned-streams|direct-source-materialization|hash-field-binding|hash-field-binding-tag)-20261103)\.json\.gz$
       - id: check-merge-conflict
       - id: check-yaml
       # BINARY_TYPES below: archive containers are binary but can START with text (a phar begins with

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -3,6 +3,8 @@
 
 # 0.11.0
 
+- Bind explicit hash-field qualifiers once while preserving dynamic defaults, copied templates and snapshot lifetimes.
+
 - Fill bounded source ranges directly into final storage and reuse owned decoder blocks during materialization, preserving read limits, errors and budgets.
 
 - Reuse caller-owned input storage for memory-backed and nested archive-member streams while preserving read errors, sparse content and shared budgets.

diff --git a/TODO.md b/TODO.md
@@ -468,8 +468,9 @@ EP/JC/P work; candidates are not speedup claims until measured on both platforms
 - [x] O22a: Qualify snapshot-owned digest observers against owning consumption, preserving cache
       insertion stability, invalidation, local-snapshot fallback, read errors and derived handoff.
       Keep [the separate mechanisms and controls](docs/benchmarks/digest-and-record-storage-20261103.md).
-- [ ] O22b: Qualify invariant field-hash binding without enlarging every unrelated template segment
-      or hiding typed state in strings. Preserve dynamic defaults and cold/non-hash controls.
+- [x] O22b: Qualify invariant field-hash binding without enlarging every unrelated template segment
+      or hiding typed state in strings. Preserve dynamic defaults and cold/non-hash controls in
+      [the binding report](docs/benchmarks/hash-field-binding-20261103.md), including rejected setup work.
 - [x] O22c1: Qualify direct plain-record construction and bulk C-escaping on both hosts, retaining
       tiny colored controls and shared display/tree consumers in the O22a report.
 - [ ] O22c2: Qualify command-owned reusable record/row buffers and streaming machine formats.

diff --git a/docs/benchmarks/hash-field-binding-20261103.json.gz b/docs/benchmarks/hash-field-binding-20261103.json.gz
diff --git a/docs/benchmarks/hash-field-binding-20261103.md b/docs/benchmarks/hash-field-binding-20261103.md
diff --git a/docs/benchmarks/hash-field-binding-tag-20261103.json.gz b/docs/benchmarks/hash-field-binding-tag-20261103.json.gz
diff --git a/docs/benchmarks/hash-field-binding-tag-20261103.md b/docs/benchmarks/hash-field-binding-tag-20261103.md
@@ -0,0 +1,59 @@
+# Hash-field binding: direct tag refinement
+
+The main-based coverage run for PR #1070 identified an obsolete hash parser whose address was
+used to classify bound fields, although rendering no longer called it. The implementation now
+uses the existing nonzero hash-binding tag and removes that parser. A segment is a field when
+it has an ordinary renderer or a hash binding; literal and unknown-field behavior, content
+counts, extraction, transforms, validation and dynamic defaults retain their contracts.
+
+This refinement is measured against the exact pushed `d95860e8e4aa5ed57e442b8b7e91b8be88fee2d2`
+implementation, separately from the [original qualification](hash-field-binding-20261103.md).
+It adds no bytes to the segment layout, public API, dependencies, locks or content caches.
+
+## Comparison
+
+Each host runs 16 in-memory cases and 20 native cases in three independent sessions, with nine
+interleaved repetitions and the fastest seven retained. Native cases discard warmups and check
+complete output and exits against independent hashlib expectations. They cover 10, 1,000 and
+10,000 files, long paths, ordinary listing, repeated fields and transformed hash fields with one
+or three requested workers. Linux uses tmpfs and CPU affinity; macOS uses warm APFS. No release
+timing overlaps builds, lint, tests or instrumentation on the same host. Negative changes mean
+less elapsed time. These are paired changes from the already optimized implementation.
+
+| Case                                                          |            Linux |            macOS |
+| :------------------------------------------------------------ | ---------------: | ---------------: |
+| Sixteen warm explicit hash fields                             |           -7.00% |           -7.98% |
+| Sixteen warm default hash fields                              |           -3.28% |           -5.71% |
+| Whole-engine repeated hash fields, 1,000 files / 1 worker     |           -3.23% |           -0.87% |
+| Whole-engine transformed hash fields, 1,000 files / 3 workers |           -1.52% |           +0.15% |
+| Ordinary name-field kernel                                    |           +0.04% |           +5.08% |
+| Ordinary component-field kernel                               |           +0.83% |           +3.32% |
+| Single warm explicit hash field                               |          +13.13% |           -1.87% |
+| Native command change range, all controls retained            | -2.01% to +3.51% | -0.55% to +0.75% |
+| Unstripped full CLI byte change                               |           -4,896 |             -208 |
+
+Small kernels have mixed costs; do not describe every row as faster. The single-field Linux
+cell and ordinary-field macOS cells remain in the lossless results with absolute times.
+Compilation changes span -1.59% to +1.35% on Linux and -6.69% to +5.27% on macOS.
+The original optimization's populations remain separate, rather than pooling timings across
+this refinement or subtracting unrelated percentage changes.
+
+## Verification and evidence
+
+The added regression distinguishes hash tags from literals, escaped hash text and unknown fields,
+and checks rendering, content counts and extraction classification. Existing hash extraction,
+all-catalog algorithms, changing defaults, copied/moved templates, snapshot lifetimes and fuzz
+cases remain active. All 313 release suites pass on each host. Nine focused ASan suites pass on each host, and nine
+MSan and nine TSan suites pass on Linux. Three explicit changed-header consumer checks pass per
+host. Focused Linux coverage of changed production field lines reaches 86/88 (97.73%) and
+67/70 branches (95.71%); this scoped check does not replace the full CI coverage gate.
+Complete release, sanitizer and explicit-header checks are recorded in the
+companion lossless archive; current-head GitHub checks remain authoritative for merge readiness.
+
+[Lossless sources, manifests, scripts, raw samples and validation logs](hash-field-binding-tag-20261103.json.gz)
+retain compiler/configuration identities, old/tag binary sizes and hashes, all successful output
+oracles and the original coverage failure. The finite-catalog overflow guard remains intact;
+coverage is fixed by deleting dead code, without excluding executable lines or weakening policy.
+
+Archive size: 585,438 bytes. SHA-256:
+`558057de3b7c3f6cb8f84632721ad0297deae2e8bd050c7edd2099ba2ef9ebd6`.
diff --git a/docs/performance-analysis.md b/docs/performance-analysis.md
@@ -3543,10 +3543,9 @@ own warmup, nine repetitions and fastest-seven selection with rotating participa
 The Linux case-isolated matrix remains separate; compare within a recorded protocol rather than
 silently treating absolute measurements from different protocols as interchangeable.
 
-O22b still needs typed field-hash binding. Compact IDs/default-use bits may fit existing segment
-padding, but actual layout, renderer-call overhead, changing context defaults and cold/non-hash
-controls must be qualified before using that space. Do not encode binding state in arbitrary
-strings or add a view-rich allocation to every segment.
+O22b's compact field-hash binding is qualified below, including layout, renderer-call overhead,
+changing defaults and cold/non-hash controls. It uses existing segment padding without encoding
+state in strings or adding a view-rich allocation to every segment.
 
 O22c2 still needs command-owned reusable output buffers and streaming rows. Keep owning/stateless
 APIs intact; a private stateful record writer could reuse capacity on the coordinator and lend a
@@ -3583,7 +3582,7 @@ in the raw artifact. The frozen formatter, source guards, independent full-outpu
 all selected root/extension release tests and focused sanitizers remain part of qualification.
 The initial nested fixture command missed `--archive-depth=2`; its warmup oracle failed before
 accepting any native row. The corrected matrix uses explicit depth two. O19b seeking/index/replay
-and admission, and O19d direct final-buffer materialization, remain open.
+and admission remain open; O19d direct final-buffer materialization is qualified below.
 
 ### O19d: fill final range and materialization storage
 
@@ -3622,3 +3621,39 @@ reserved storage directly; its expected owned-read counter changes from one to z
 parent lifetimes, content and shared-budget tests remain. No new public flag, external dependency,
 worker, lock, metadata probe or host-I/O bypass is introduced. O19b seeking/index/replay/admission
 remains open.
+
+### O22b: bind hash fields without growing unrelated segments
+
+A private two-byte binding fits beside the existing post-processing enum. Algorithm IDs derive
+from the existing `hash::AlgorithmNames` catalog; encoding choices distinguish explicit values
+from context defaults. This avoids qualifier parsing and default work for explicit choices while
+preserving changed defaults on every render, copies/moves, invalid specifications, post-render
+transforms and snapshot ownership/invalidation. No borrowed qualifier view is retained.
+
+Compilation validates the binding once. Bare/transformed defaults bind directly, and rendering
+uses the existing nonzero binding tag to classify hash fields; ordinary renderers retain their direct calls.
+A coverage-driven refinement removes an obsolete parser that had served only as a function-identity
+tag. The first prototype duplicated static
+validation and read a separate tag for every field. It increased hash compilation costs by about
+30-50% and is not selected; complete first/refined populations remain separate.
+
+[Cross-platform qualification](benchmarks/hash-field-binding-20261103.md) includes warm/cold
+explicit, partial and default specs, compilation, ordinary fields, whole-engine in-memory runs and
+native listing/hash-field commands. Explicit cached kernels improve about 41-56% in independent
+confirmation; larger in-memory commands improve about 2-5%. Native gains are smaller. The Linux
+simple-name field kernel retains about a 2% cost, roughly one nanosecond per entry; cold and Mac
+controls vary across populations. Preserve their absolute times and identical-baseline movement;
+this is not a universal rendering speedup.
+
+Tracked template storage is unchanged across measured shapes. This is conservative capacity
+accounting, not an allocator or RSS claim. Full CLI footprint increases 1,440 bytes on Linux and
+368 bytes on macOS. Three new tests cover changing defaults across copies/moves, all shared-catalog
+algorithms and invalid dynamic defaults with independent explicit choices. Existing derivation,
+extraction, snapshot, output and fuzz suites remain; complete release and focused sanitizer
+qualification passes both hosts. O22c2 command-owned buffers and streaming rows remain open.
+
+The [direct-tag confirmation](benchmarks/hash-field-binding-tag-20261103.md) retains old/tag sources
+and three-session cross-platform measurements. Sixteen warm explicit fields improve another 7-8%,
+while small ordinary kernels and one Linux single-hash cell retain costs. Native differences remain
+small, and the full CLI decreases 4,896 bytes on Linux and 208 bytes on macOS relative to the prior
+binding implementation. A structural regression preserves literal/hash/extraction distinctions.
diff --git a/release-site.json b/release-site.json
@@ -80,7 +80,9 @@
     "docs/benchmarks/archive-input-buffers-20261102.md": "docs/benchmarks/archive-input-buffers-20261102.html",
     "docs/benchmarks/digest-and-record-storage-20261103.md": "docs/benchmarks/digest-and-record-storage-20261103.html",
     "docs/benchmarks/caller-owned-streams-20261103.md": "docs/benchmarks/caller-owned-streams-20261103.html",
-    "docs/benchmarks/direct-source-materialization-20261103.md": "docs/benchmarks/direct-source-materialization-20261103.html"
+    "docs/benchmarks/direct-source-materialization-20261103.md": "docs/benchmarks/direct-source-materialization-20261103.html",
+    "docs/benchmarks/hash-field-binding-20261103.md": "docs/benchmarks/hash-field-binding-20261103.html",
+    "docs/benchmarks/hash-field-binding-tag-20261103.md": "docs/benchmarks/hash-field-binding-tag-20261103.html"
   },
   "generated_html": {
     "XFF.md": {