This patch reduces allocations and speeds up memory analysis, numeric matrices, text handling and fixed-record
operations. It also reduces the NuGet package size and adds reproducible performance comparisons.

Changed

• The NuGet package's CStructSharp.xml documents only the public API. The compiler also wrote the documentation of
  every internal and private member, about three quarters of the file, which IntelliSense never shows. Each copy
  shrinks from 2.1 MB to 0.5 MB and the package from 1.68 MB to 1.15 MB. No public documentation is removed, and
  package validation now rejects internal documentation.

Performance

• Reduce memory-session member-selection allocations for wide structs and share equivalent built-in scalar
  preparation within each memory schema. Preserve promoted-member failures, source interactions and per-type
  validation; caller-supplied codecs and declarations keep independent preparation.
• Materialize memory-session numeric arrays directly into owned storage while preserving element-by-element source
  requests, cancellation and budgets. Bound BTF name lookup caching and share successful scalar/bit-slice compilation
  within each metadata table, keeping independent root graphs, diagnostics and validation.
• Reduce temporary allocations when serializing bounded text and short wide-character fields, including runtime
  span, stream and async output. Runtime writes retain normal tiered compilation to limit first-use cost.
  Generated bounded-text writes encode directly into their destination after validation.
• Reduce temporary storage and per-character work when runtime readers decode wide and byte-counted text from
  memory, including buffered async input, while preserving owned strings and original failure behavior.
• Materialize runtime numeric matrix rows without a temporary flat reference array, and reduce allocation while
  validating matrices for serialization. Specialize generated fixed bitfield leaf readers while preserving repeated
  read charges, caller construction effects and original failure paths.
• Bulk-encode generated numeric matrix rows into owned serialization output, retaining scalar writes for borrowed
  spans and failure fallbacks. Construct complete nested runtime values inside general-engine static plans without
  repeated presence bookkeeping; preserve captures, mutable ownership and cancellation checkpoints.
• Decode generated numeric multidimensional arrays directly into owned rows, removing flat temporary arrays and
  copying. Reduce generated UTF/Latin-1 text scratch storage and padded Latin-1 string allocations while preserving
  strict validation and error details.
• Speed up fixed-root runtime reads by constructing complete member slots, reduce runtime numeric call overhead,
  and inline bounded generated fixed-record helpers. Preserve validation, conversions, ordering, mutation and
  owned results; retain general-path fallbacks.

Measured against 0.11.1. These results include the changes shipped in 0.12.0 as well as this patch; the
0.12.0 migration notes still apply. Identical benchmark sources, inputs and complete owned results were measured
on an AMD Ryzen 9 9950X, Windows 11, .NET 10.0.12, with production tiering and PGO enabled. Three alternating
process pairs used 24 warmup iterations and 15 measured iterations each. Times are medians of launch medians;
speedup ranges are the observed paired ratios, not confidence intervals. Every measured iteration was retained.

These eight cases showed repeatable reductions under the comparison's screening rules. Of the other 22 selected
cases, five stayed within the timing margin and 17 remained unstable. In particular, small generated serialization,
mapped packet reads and 256 generated text records do not support a speedup claim. Wide-member selection
still reduced allocation from 26,400 to 288 bytes, and preparation of 1,000 pointer definitions from 18.7–19.0 MB
to about 0.515 MB, but their timings remained variable. The initial six-warmup campaign also remains inconclusive.
See the complete measurements, limitations and reproduction commands,
including every selected case and the retained samples. These observations do not establish performance on other
workloads or machines.

Complete memory-image workflows. The application sources are identical on both versions. Three fresh process
pairs per runtime use one private Linux memory image with a warm operating-system cache. Bootstrap and first calls
include preparation; repeated calls start after five seconds of warmup. All workflow-result and symbol hashes match
across all 12 JIT/Native AOT processes. These are capture-specific results, not cold-disk throughput measurements.

Documentation and tooling

• Refresh the README serializer comparison and add a measured comparison with 0.11.1, including allocation costs,
  production JIT and Native AOT results, uncertainty, and reproduction instructions.
• Refresh the explorer catalog for 73 added regression tests. Re-review mutation-equivalence entries after codec
  changes and remove four proofs that no longer apply; test and mutation scopes remain unchanged.
• Add a fast managed development test profile, bounded parallel test orchestration and parallel generator tests.
  Keep exhaustive sweeps in full, CI and release runs, and document the edit-loop and final-validation commands.
• Document the measured speed and reliability limits of short Impact comparisons, with isolated runners that reuse
  existing benchmark cases, including a production tiering/PGO comparison runner. Clarify that 3% flags and
  smallest-median selection do not establish confidence.
• Add development performance comparisons with immutable, source-validated build bundles, automatic CPU affinity,
  full-sample timing and allocation reports, and targeted confirmation. Document the workflow for contributors and agents.

Full changelog

What's Changed

• build(deps-dev): bump cspell from 10.3.4 to 10.3.5 in /docs in the documentation-dependencies group by @dependabot[bot] in #49

Full Changelog: v0.11.1...v0.12.0

Full Changelog: v0.11.0...v0.11.1

Full Changelog: v0.10.0...v0.11.0

Full Changelog: v0.9.1...v0.10.0

Full Changelog: v0.9.0...v0.9.1

Full Changelog: v0.8.3...v0.9.0

Full Changelog: v0.8.2...v0.8.3

What's Changed

• build(deps): bump file-type from 22.1.0 to 22.1.1 in /apps/inspector by @dependabot[bot] in #27
• build(deps-dev): bump @types/node from 26.6.1 to 26.6.2 in /apps/inspector by @dependabot[bot] in #28
• build(deps-dev): bump markdownlint-cli2 from 0.23.2 to 0.23.3 in /docs by @dependabot[bot] in #30
• build(deps): bump the frontend-dependencies group across 1 directory with 4 updates by @dependabot[bot] in #33
• build(deps): bump vue from 3.5.42 to 3.5.43 in /apps/inspector by @dependabot[bot] in #31
• build(deps): bump @iconify-icons/vscode-icons from 2.0.4 to 2.0.6 in /apps/inspector by @dependabot[bot] in #29
• build(deps-dev): bump eslint from 10.10.0 to 10.11.0 in /apps/inspector by @dependabot[bot] in #32

Full Changelog: v0.8.1...v0.8.2

Full Changelog: v0.8.0...v0.8.1