Reduce native scalar admission allocations

Author: KSemenenko

AGENTS.md

@@ -393,6 +393,9 @@ For changes outside existing owner authorization, obtain direction before changi
 ## Preferences
 
 ### Likes
+- Each agent MUST follow the execution plan for its own user-assigned task and complete only that task; change its scope only when the owner explicitly redirects it (owner correction 2026-10-03).
+- Other chats' plans, progress, requests and concurrent changes MUST NOT expand an agent's task or redirect its implementation. Preserve their work and leave their tasks to their owners (owner correction 2026-10-03).
+- Repairs and verification MUST stay within the assigned task. Report an unrelated failure to the owner without taking over its implementation; direct user instructions remain authoritative (owner correction 2026-10-03).
 - Frequent concise Ukrainian progress stating completed work, remaining work and real blockers.
 - The root README MUST introduce KeyLoad as a product: its purpose, connected data models, agent use cases, getting started, and direct credits to the projects it actually uses. Keep dependency versions in their canonical manifests and link to detailed technical documentation (owner correction 2026-10-03).
 - Shared-checkout work that preserves unrelated changes and completes authorized delivery.

docs/ADR/ADR-060-native-internal-serialization.md

@@ -437,3 +437,36 @@ original report inputs, exact ratios/intervals/launch summaries, source diff and
 independent reviewer signature. Controlled arithmetic cannot authenticate GitHub.
 Local BDN experiments are owner-authorized development evidence with actual
 source/machine/settings, and cannot qualify website/cluster results.
+
+
+## NSP006 development verification and assigned scope
+
+AC-IS-PERF005 measurement scheduling is explicitly isolated after the actual full
+v7 report measured Normalize262480B and Require262144B while the focused exact
+comparison passed. This demonstrates a context-dependent measurement difference;
+it does not prove a specific CLR cache/GC cause or extra production allocation.
+The two scalar allocation tests use the pinned TUnit1.72.10 no-key NotInParallel
+class constraint, whose actual package XML specifies exclusion from every other
+test. Preserve the same32 warmups,4096 iterations, synchronous current-thread
+windows, outside-window data/assertions, exact0B and exact equality. No tolerated
+bytes, min-of-retries, extra cache, GC manipulation, removed case or changed
+production code. This follows the existing JsonTextAllocationTests scheduling
+boundary; all other tests remain parallel. The scalar worker owns only the one
+class scheduling attribute after root acceptance; root verifies the full normal/
+scalar suite and retains the original failed report.
+
+The latest owner instruction confines this delivery to native Orleans serialization
+latency and allocations: NativeWireSupported and its four scalar regression files.
+The R19 process-observation repair branch was cancelled; all eight exact task-owned
+source edits were undone, with original sources and reports retained outside the
+checkout. The existing process-lifetime failure remains reported as unrelated
+verification evidence. No weakened assertion or process-helper change is delivered.
+
+The historical native diagnostic run37129441460 at3ae408fe remains authenticated.
+Do not dispatch its former internal mode through benchmarks.yml; that workflow
+owns end-to-end database comparisons. Local full24-cell BDN pairs retain exact
+corpora, profile, actual host, controls and original samples. A source image must
+retain every tracked base file and only its declared overlays. Its results prove
+only that image, not a green shared checkout or delivered GitHub qualification.
+Images containing cancelled R19 overlays remain historical evidence and cannot
+qualify the final five-source delivery. Required exact-source CI/RF3 gates remain.

docs/Features/BenchmarkComparisons/NativeSerialization.md

@@ -70,4 +70,38 @@ reaches actual API subprocesses and waits. Pure step/deadline regression data
 never authenticates GitHub or qualifies results. A new exact-source job must
 complete all original report/corpus/generated/provenance/artifact checks.
 
-REQ/AC-IS-PERF005..007 now have the accepted exact terminal-allocation/admission and prospective quantitative/variance contract in ADR060 (NSP006). Test matrix: new NativeWireSupportedScalarTests/Fixtures prove frozen allowed terminals and generic-owner/open closures; ScalarAllocationTests prove measured0B auxiliary delta, nullable Normalize equality; ScalarDepthTests prove264/265 boundaries. Retain all existing native roundtrip/fault/resource tests. AC007 manual evidence exception: actual authenticated before/after originals, corpus/cohort checks, exact interval/ratio/launch arithmetic and independent review. Protected baseline gate remains pending R18 fresh artifact.
+REQ/AC-IS-PERF005..007 now have the accepted exact terminal-allocation/admission and prospective quantitative/variance contract in ADR060 (NSP006). Test matrix: new NativeWireSupportedScalarTests/Fixtures prove frozen allowed terminals and generic-owner/open closures; ScalarAllocationTests prove measured0B auxiliary delta, nullable Normalize equality; ScalarDepthTests prove264/265 boundaries. Retain all existing native roundtrip/fault/resource tests. AC007 manual evidence exception: actual authenticated before/after originals, corpus/cohort checks, exact interval/ratio/launch arithmetic and independent review. The protected R18 baseline gate is admitted by the complete original artifact from run37129441460 at3ae408fe; see the current delivery record below.
+
+## Current delivery and workflow contract, 2026-10-03
+
+The later explicit owner workflow correction supersedes the historical
+AC-IS-PERF-003/004 placement and dispatch requirements above: benchmarks.yml is
+reserved for end-to-end database comparisons. Internal native/raw/codec
+microbenchmarks cannot be its jobs, modes or dependencies. Do not repeat the old
+native-only dispatch. The former job and its unchanged original artifacts remain
+historical evidence; removing its pipeline surface must preserve the actual
+fixture, correctness tests and database comparison gates. Local complete paired
+BDN runs use these same fixtures, six corpora,24 cells and2/3/6/200ms profile,
+with actual source/host/children/exporters and explicit development-only status.
+They do not refresh website performance evidence.
+
+NSP005 baseline admission is complete: exact3ae408fe CI run37129421675 passed
+normal/scalar2637 each, recovery194, RF3 SDK/MCP67 and analyzers118 with no
+skips. Protected run37129441460 retained24 cells, six corpora,48 distinct real
+children and matching127-case normal/scalar diagnostic suites. Artifact11276836758
+has verified SHA256 b3353fa730b664cb70c84dfb49ad9d4f0827849b9e15684998c37f7ce783ec83;
+independent source joins and credential review passed. These results qualify
+only the authentic3ae baseline, never a later candidate.
+
+NSP006 writes are accepted under ADR060 AC-IS-PERF005/006/007. The sole runtime
+change fast-returns existing closed nongeneric terminal types after one unchanged
+Normalize call; the complete fallback and persisted/native corpus bytes remain
+unchanged. Root integrates only the production change and four scalar regressions;
+the owner cancelled the unrelated process-helper branch. The complete local24cell
+pair passed all12native allocation budgets, with Command NativeDecode64.62%/79.11%
+lower allocations at1024/16384. Only3of12JSONmean controls passed; AC007 latency
+qualification remains open. These per-operation allocations do not establish total
+RAM usage or global maximum speed. Final five-source normal/scalar/recovery and
+delivered-source Linux RF3 CI remain distinct from historical image evidence.
+Current original receipts, failures and scope limitations live in
+[stage004](../../implementation/native-serialization-repair-stage-004.json).

docs/Features/InternalSerialization.md

@@ -63,3 +63,36 @@ diagnostic job do not qualify runtime, acceleration, RF3 or durability.
 
 
 R17 under REQ-IS002/AC-IS002 and AC-IS004/007 distinguishes persisted native queue-body corruption from caller validation, preserving bytes, ready state, counters, positions and absent outcome before successful restore with the same request identity. QueueBodyAccountingTests/QueueBodyFailureAssertions own this real-ZoneTree fixture. Unknown well-known header metadata is classified only at the official single-header/span boundary in NativeFieldHeaderReader, using the saved official Reader and registered type table; arbitrary owned/generated/session key failures remain visible. NativeUnknownWellKnownHeader* tests use genuine writer twins, extended field deltas and public/MCP controls. ADR060 owns the accepted implementation and rollback. Normal/scalar/recovery/RF3 qualification remains exact-source GitHub-only; source review and a development build cannot satisfy it.
+
+## Terminal admission allocation reduction
+
+REQ-IS-PERF005 / AC-IS-PERF005 requires zero auxiliary allocation for warmed,
+normalized terminal types and nullable admission allocation equal to the existing
+Normalize loop. REQ-IS-PERF006 / AC-IS-PERF006 preserves the supported domain,
+closed generic-owner enum fallback and exact type-depth264/265 boundary. Only
+NativeWireSupported's existing primitive/enum/scalar predicate receives a root
+shortcut; the complete iterative fallback, generated codecs, persisted bytes,
+resource limits and validation remain unchanged. NativeWireSupportedScalar
+AllocationTests, Tests, Fixtures and DepthTests own the positive, unsupported,
+nullable, byte-ownership and depth controls. ADR060 records rollout and rollback.
+
+REQ-IS-PERF007 / AC-IS-PERF007 uses the prospective allocation/latency budgets,
+complete24-cell profile, six exact corpora, twelve JSON cohort controls and
+conservative interval decision in ADR060. Pair arithmetic requires retained
+originals and independent review; two local command-decode cells cannot satisfy
+this full criterion. Internal microbenchmarks follow the latest root policy and
+must not be dispatched through the end-to-end database comparison workflow.
+
+The [stage004 evidence](../implementation/native-serialization-repair-stage-004.json)
+binds the complete successful normal/scalar/recovery/analyzer/RF3 CI and protected
+24-cell diagnostic baseline to delivered commit3ae408fe. The pending shortcut's
+full local24cell experiment preserves all six corpora and native/JSON bytes.
+Command NativeDecode allocation falls from128270 to45386B/op for1KiB and from
+1541554 to321975B/op for16KiB (64.62%/79.11%). All12native allocation budgets
+pass. Only3of12JSONmean controls pass, so the pair does not qualify latency or
+AC-IS-PERF007. These are per-operation allocations, not total server RAM.
+Final five-source build, normal/scalar, recovery and exact-source RF3 gates
+remain separate from the historical image that produced this pair.
+The owner cancelled the unrelated process-helper repair branch. This delivery
+contains only the admission shortcut and its four scalar regression files;
+source-image results containing cancelled helpers remain historical evidence.

docs/Features/InternalSerialization/Acceptance.md

@@ -42,3 +42,11 @@ System/derived-collection native codec shapes must fail before allocation.
 Bounded unknown fields, known-schema references and the actual owned generated
 closure retain positive coverage. Full arbitrary native codec compatibility is
 out of scope and requires a separately accepted bounded shape contract.
+
+## NSP006 criteria
+
+| ID | Pass/fail contract | Automated evidence |
+|---|---|---|
+| AC-IS-PERF005 | Warm normalized terminal admission allocates0 auxiliary bytes; nullable admission equals independent unchanged Normalize allocation. No caches or changed limits. | NativeWireSupportedScalarAllocationTests use the no-key TUnit NotInParallel constraint; exact byte assertions,32 warmups/4096 iterations and outside-window data remain unchanged. |
+| AC-IS-PERF006 | Supported closed terminal roots pass; open/unsupported roots and unsupported generic-owner arguments fail with the original corruption error. Fixed263 array wrappers pass;264 fail. Generated nullable values/nulls, ordered collections and independent bytes survive round trips. | NativeWireSupportedScalarTests/Fixtures/DepthTests plus the complete existing native normal/scalar suite. |
+| AC-IS-PERF007 | All prospective budgets, six exact corpus receipts, full24-cell profile, twelve JSON controls and confidence/cohort conditions in ADR060 hold. Missing/drifted/indeterminate pairs fail qualification. | Actual complete BDN originals and strict report validation; explicit manual arithmetic exception requires source/host/profile-bound originals and independent review. Local two-cell data is insufficient. |