From 9e61f158c0283f09fd125cb9e4d38c3d06a608f7 Mon Sep 17 00:00:00 2001 From: Thorsten Fuchs <76904008+ThorFuchs@users.noreply.github.com> Date: Mon, 12 Oct 2026 11:31:00 +0200 Subject: [PATCH] Document measured integration review and precise release blockers --- .../Release1.1/IntegratedPerformance.md | 55 +++++++++++++++++++ .../Architecture/Release1.1/Mobile.md | 7 +++ .../Architecture/Release1.1/Progress.json | 20 +++++++ .../Architecture/Release1.1/Reading.md | 7 +++ .../Architecture/Release1.1/Review.md | 35 ++++++++++++ 5 files changed, 124 insertions(+) create mode 100644 Documentation/Architecture/Release1.1/IntegratedPerformance.md create mode 100644 Documentation/Architecture/Release1.1/Progress.json create mode 100644 Documentation/Architecture/Release1.1/Review.md diff --git a/Documentation/Architecture/Release1.1/IntegratedPerformance.md b/Documentation/Architecture/Release1.1/IntegratedPerformance.md new file mode 100644 index 0000000..62d704b --- /dev/null +++ b/Documentation/Architecture/Release1.1/IntegratedPerformance.md @@ -0,0 +1,55 @@ +# Integrated 1.1 performance disposition + +The unchanged twelve-workload public harness compares the released 1.0 core at +`a2733dd` (identical Sources/benchmark harness to tag 1.0.0) with the 1.1 core plus +capacity-history correction. M1 Max/32 GiB, macOS27.0, Swift6.4, Xcode27.2/SDK27.2, +release optimization, finite balanced profile, three warmups, nine samples and +three invocations/sample apply to both. Inputs/configuration hashes match exactly. +Measurements are serial, with compilation/safety workloads idle. + +An initial implementation recorded maximum admission history on every meter check +and regressed most workloads substantially. Retain those results. The correction +records history only in native coding/read scopes that can tighten limits; work is +already cumulative and live retention uses current ownership. Fixed-policy codecs +skip unused history writes. Refusal/cancellation/overflow remains terminal. The +capacity-history refusal test protects against silently weakening later ceilings. + +| Workload | Before / after ms | Change | Before / after logical peak bytes | +| --- | ---: | ---: | ---: | +| parse_small_exact | 0.018653 / 0.020917 | +12.14% | 649 / 721 | +| parse_owned_data | 0.023042 / 0.025722 | +11.63% | 0 / 0 | +| parse_large_exact_array | 21.826889 / 22.917903 | +5.00% | 320064 / 320064 | +| write_large_unicode_string | 100.354583 / 101.423028 | +1.06% | 1040004 / 1040004 | +| native_value_codable | 16.128958 / 16.260583 | +0.82% | 166332 / 166468 | +| exact_number_parse_compare | 1.459167 / 1.448500 | -0.73% | 0 / 0 | +| schema_compile_repeated_patterns | 36.223806 / 35.530042 | -1.92% | 103074 / 104642 | +| shared_plan_regex_evaluation | 0.352625 / 0.358167 | +1.57% | 1412 / 1412 | +| schema_exact_evaluation | 0.022833 / 0.023208 | +1.64% | 647 / 647 | +| staged_rfc6902 | 1.522097 / 1.531833 | +0.64% | 97100 / 97100 | +| rfc7396 | 0.052417 / 0.054500 | +3.97% | 1191 / 1255 | +| generated_root_array_stream | 1179.382305 / 1365.204694 | +15.76% | 52 / 52 | + +The unchanged 10% review trigger applies. Accept the observed small exact parse +(+12.14%, about 2.26 microseconds) and owned-Data parse (+11.63%, about 2.68 +microseconds) costs, including controlled duplicate/index bookkeeping and terminal +accounting. Small-parse logical peak rises 649 to721 (+11.09%): explicit additional +index capacity is covered by unchanged weights/limits. Accept the stream observation +(+15.76%, about1.86 microseconds/record) for this correctness/ownership increment, +with no throughput improvement claim or isolated causal attribution. Raw samples +and variance remain material evidence. No profile/10% threshold is raised. + +The historical stream 1.0749626803 versus0.968493986 seconds (+10.9932%) retains +its earlier disposition independently. The new matched stream pair is1.3652046943 +versus1.1793823053 seconds; do not compare either new run with a historical median +as if inputs/environment were a fresh matched pair. + +The separate wide-index construction/lookup corpus retains its full trigger review +and spare-capacity tradeoff in IndexMeasurements.md. This integrated corpus supplies +no new index crossover, SIMD/parallel speed claim or SwiftyJSON comparison. + +Physical iPhone14ProMax/iOS27.2 current-candidate public-product/payload work has +nine release samples after three warmups: median54.406542 ms and process peak RSS +24395776 bytes. This is a whole app/SDK peak, including UIKit, without isolated +allocation or hard-RSS claims, and no matched mobile optimization comparison. +Physical iPad measurements remain blocked. Logical bytes, allocator observations, +process RSS and caller computation remain distinct. diff --git a/Documentation/Architecture/Release1.1/Mobile.md b/Documentation/Architecture/Release1.1/Mobile.md index f692b5b..00816a3 100644 --- a/Documentation/Architecture/Release1.1/Mobile.md +++ b/Documentation/Architecture/Release1.1/Mobile.md @@ -27,3 +27,10 @@ Swift Testing and isolated iPhone/iPad iOS 27 simulators execute three-process disk save/reopen/update checks. The physical iPhone is connected with iOS 27.2 and developer mode enabled; signing/runtime/performance are separately qualified. The physical iPad is unavailable and remains a mandatory R5 release blocker. + +A development-signed qualification wrapper now executes the exact public harness +on physical iPhone14ProMax/iOS27.2. Public smoke, eight fixtures, three separate app +processes for disk save/reopen/update and nine release measurements are materialized +as run-ID-bound files retrieved from its own container. The temporary app is +removed afterward. Core consumers/source remain independent of UIKit/signing; +no public UI product or server semantics are added. Physical iPad remains blocked. diff --git a/Documentation/Architecture/Release1.1/Progress.json b/Documentation/Architecture/Release1.1/Progress.json new file mode 100644 index 0000000..17fc5f8 --- /dev/null +++ b/Documentation/Architecture/Release1.1/Progress.json @@ -0,0 +1,20 @@ +{ + "schema_version": 1, + "target_release": "1.1.0", + "milestone": "implementation, runtime qualification and reviewable candidate", + "tested_core_sha": "a48eb337b70eb075ce37c99dc3d150032f720d00", + "preserved_history": ["1.0.0", "W00-W12", "G01-G12", "RC01-RC08", "F01-F06", "J01-J04", "R0 turn 001"], + "root_function_counts": {"JSON":86,"JSONStreaming":11,"JSONSchema":38,"JSONRegexRuntime":12,"JSONPatch":9}, + "root_function_total":156, + "parameterized_invocations":173, + "cap_probes":49, + "readme_examples":10, + "authored_public_docc_symbols":582, + "payload_fixtures":8, + "r5_matrix": {"macOS":"executed","iPhone_simulator":"executed","iPad_simulator":"executed","physical_iPhone":"executed","physical_iPad":"blocked; unavailable"}, + "conditional_features": {"SIMD":"deferred","parallel_batches":"deferred","E11":"deferred","E12":"deferred"}, + "evidence_policy":"Actual final native/safety/docs/mobile receipts and materialized Turn002 archive identify commands, source hashes, counts and limitations. Documentation-only successors map unchanged core/test inputs explicitly.", + "publication_performed":false, + "remaining_release_blockers":["Physical iPad smoke, disk persistence and release measurements for the agreed R5 matrix", "Complete maintainer review and protected PR integration before final release readiness"], + "performance_review":"IntegratedPerformance.md; unchanged 10% trigger, retain initial regression measurements and final accepted dispositions" +} diff --git a/Documentation/Architecture/Release1.1/Reading.md b/Documentation/Architecture/Release1.1/Reading.md index 173d029..23808e7 100644 --- a/Documentation/Architecture/Release1.1/Reading.md +++ b/Documentation/Architecture/Release1.1/Reading.md @@ -50,3 +50,10 @@ select year/month/day plus fractional UTC time. Baseline epoch-only tests could detect losing the calendar date. Record this as a real affected coding qualification failure and refresh its native/safety/consumer/DocC evidence; do not redefine it as permitted rounding or silently reduce the assertion. + +Native coding and read scopes explicitly enable accepted-capacity history because +projection/payload limits may tighten after prior admission. Fixed-policy operations +retain terminal health/work accounting but skip history they cannot consume. Work +counters and live ownership supply their own restriction checks; no meter copy or +quota reset is introduced. The final integrated performance review records this +measured correction and retained regressions. diff --git a/Documentation/Architecture/Release1.1/Review.md b/Documentation/Architecture/Release1.1/Review.md new file mode 100644 index 0000000..7ea7982 --- /dev/null +++ b/Documentation/Architecture/Release1.1/Review.md @@ -0,0 +1,35 @@ +# Separate candidate review pass + +Reviewer: Codex, the implementing agent, in a separate static/evidence review pass +on 2026-10-12. This is not an independent human review or a GitHub approval. +Reviewed core: a48eb337b70eb075ce37c99dc3d150032f720d00 against main a2733dd. +Documentation successors retain those core bytes. Read the actual source, compiler +receipts, test bodies, SDK/runtime results and raw measurements; no summary-only audit. + +| Finding | Severity | Resolution and evidence | +| --- | --- | --- | +| Native ISO-8601 omitted calendar fields | high | Explicit year/month/day UTC style in both directions; exact modern-date regression plus original coding cases | +| Base64 accepted nonzero unused padding bits | medium | Shared strict decoder rejects AB==/AAB= and noncanonical inner-null spellings | +| Caught owned failure could certify callback completion | high | Terminal meter health and final checkpoints; caught numeric/payload/cancellation/retired-counter overflow cases | +| Simulator bundle had unsuitable Resources layout | high | Direct Draft2020-12 resource root; default-engine signing, integrity loader and both SDK/runtime consumers | +| SDK retained a container beyond lexical scope | qualification | Preserve failed simulator weak-reference proof; three separate exited processes certify true disk reopening on Mac/simulators/physical iPhone | +| Exact CI count labels were mismatched | qualification | Inspect actual bundles; exactly JSON86/Streaming11/Schema38/Regex12/Patch9, 173 parameterized invocations; retain initial failures | +| Initial capacity-history writes regressed broad workloads | performance | Enable history only where later restriction is supported, exclude existing work/live counters; retain initial/final matched samples and every 10% disposition | +| Physical launch gave no inspectable results | qualification | Explicit UIKit entry/scene lifecycle and run-ID-bound result files retrieved from its own app container; normal child exits verified | +| Benchmark absolute output appended to cwd | tooling | Byte-equal relocation/path map; use the documented relative output thereafter, retain original commands | + +Inspect every object publication route, exact collision confirmation, pre-growth +spare/overlap reservations, cancellation within loops and nonthrowing finish. +Review cursor owner checks before mutable access, source-map parent spans, complete +bounded paths, scope invalidation/root release, numeric fidelity and the single +projection ledger. Review payload validated-bytes-only storage, complete grammar/EOF, +balanced intersections, checked expansion, corruption versus operational errors +and absence versus null. Native/foreign callbacks hold no active inout meter borrow. +No blanket unsafe/actor annotations, permissive parser or regex fallback is added. + +Retain 1.0/W00–W12/G01–G12/RC/F/J and R0 history. Native/safety/docs, independent +policies, all49 caps, public consumers and physical/simulator receipts identify +actual commands/source and nonzero counts. Source archives exclude private input, +Git/build trees, signing profiles and device identifiers. Physical iPad remains +blocked; conditional SIMD/parallel batches stay deferred, with no unqualified +performance claim. No merge/tag/publication is implied by this review.