Eve library

V2 rollback native probe · 2026-09-07

Guide

Revision 1

On this page

This receipt establishes a reproducible native-build boundary for the next rollback explanations. No rollback runtime test passed or executed. No new slide, teaching assignment, narration or product-source repair is included. The complete presentation objective remains open.

Observed results#

Check Actual result What it establishes
Fresh unchanged-pin preparation and integrity check 319 source files, four host templates; pass Inputs agree with Git and templates, not runtime behavior
Unchanged-pin action discovery Exit 6; two UnrealHeaderTool errors The inspected local host cannot generate these Blueprint-exposed hash fields
Explicit diagnostic fixture Two BlueprintReadWrite annotations removed; all other source bytes unchanged A qualified build experiment, not a passing baseline or Blueprint repair
Diagnostic action discovery Exit 0; 296 reflection outputs reported, 95 planned actions Header generation and a project-local action plan, not completed compilation
Diagnostic module build Seven V2Input compiler errors; explicitly stopped, exit 143 Additional dependency blockers; not a naturally completed build result
Test translation unit alone Compiler exit 0; object contains all five RunTest definitions Harness compilation only; no linked test module or executed assertions
Preparation/integrity regressions 18 pass Copy fidelity, adaptation limits, drift rejection and output safety
Complete presentation unit suites 205 Python and 274 Node tests pass Existing tool/content regressions plus the new setup checks; not C++ runtime coverage

state.json binds the source manifests, logs, action plan, compiler response files, engine/tool identities and checks. Its explicit false runtime, release and whole-objective fields must remain false. The original pre-tool- hardening header failure is retained separately from the fresh baseline run.

Source and engine boundaries#

The source revision is 1ceae52c7896a59ac74a4429d410f3218f0b91c6, independently of the working branch. The preparation tool copies V2Core, V2Input, V2Gameplay, V2Combat, V2Netcode and the V2AdaptiveAI plugin. The test host does not reproduce the full V2 target's module inventory, strict-FP overrides, pre-build code generation or authored assets.

The local installation's Build.version identifies Unreal 5.5.4. Build commands ran as ueagent using its installed UBT and bundled Clang 18 toolchain. The engine checkout has three pre-existing tracked modifications, recorded in the state; no engine-source edits were made here. Binary and selected engine-source hashes identify the tools actually used. There was no editor runtime launch.

Raw .txt records retain exact tool-produced bytes, including trailing spaces and the engine version file's final blank line. Receipt-local Git attributes disable newline conversion and whitespace-style checks for these records only; authored code and prose retain their normal checks.

At V2NetcodeTypes.h:429 and :478, UHT rejects Blueprint-visible uint32 members named FrameInputHash and FinalInputHash. The corresponding local UHT rule checks IsMemberSupportedByBlueprint; UInt32 does not add that capability. The diagnostic fixture removes just those two Blueprint visibility flags. Exact replacement guards and original/copied hashes prevent this from silently changing field types or C++ function bodies.

The diagnostic compile then reports:

  • V2InputBlueprintLibrary.cpp:451: mixed int32/uint16 arguments to FMath::Min.
  • V2InputBlueprintLibrary.cpp:806, :809, :812, :815: no EKeys::Touch1 through Touch4 in this engine.
  • V2InputTypes.cpp:3144: NewDodgeTimingAssistMode is undeclared; the method parameter is NewDodgeTimingMode.
  • V2InputTypes.cpp:3358: FString does not match the inspected helper's const TCHAR* argument.

No additional diagnostic or product fixes were applied. Stopping after these errors avoids turning a presentation verification attempt into an unreviewed product repair. It leaves dependency linking and native execution genuinely unverified, not a reason to mark the overall presentation goal blocked or done.

Test and process evidence#

The five source-derived candidates cover prediction beyond the stored horizon, correction followed by a new live frame, recovery requiring a successful correction, resend/ACK boundaries and budget/restore validation limits. They compile against the diagnostic host's generated headers with the exact UBT-produced test response file. Its nested response, forced definitions, precompiled-header identity and resulting object identity are retained or hashed. The object itself is not distributed. A symbol listing confirms five RunTest definitions; that does not mean the automation framework discovered or invoked them.

Preparation manifests always record compiled: false and testsExecuted: false; they are input receipts, not mutable success flags. The tests deliberately reject altered source even when its manifest is self-updated, wrong fixture mode, changed/missing templates, changed modes, additional source files, symlinks and nonempty output directories. An initial test's expected annotation line numbers were corrected from 428/477 to the observed 429/478 before the passing run; the source was not changed to satisfy that expectation.

Commands used one compiler action at a time. The safety wrapper sampled host resources every 25 seconds and delegated owned-process-group cleanup to the repository supervisor. The module build was explicitly terminated after errors; the observed wrapper exit was 143, and no owned compiler/build process remained. The first compiler command's -NoUBA did not disable the separate local UBA executor, and its -StopAfterErrors was not the inspected executor's supported fail-fast setting. Preserve those command/log facts. The fresh baseline recipe adds -NoUBALocal; future builds must use supported fail-fast configuration or explicit error supervision. Resource safety was maintained; no broad workspace build, engine rebuild, browser or narration job ran alongside compilation.

Delivery continuity and remaining work#

The builder and coverage/inventory freshness checks are recorded in the state. All 85 delivered artifacts and 409 recording hashes from the contracts receipt are checked unchanged. No new browser, audio decode, listening or visual pass is claimed: this increment changes verification setup and source-review records, not presentation UI or media. The prior browser/PDF/audio evidence retains its original scope. Source/teaching completion remains pending.

Continue the rollback learning map: resolve a faithful native execution route with every adaptation disclosed, execute and extend the five candidates, finish packet/fallback/snapshot and remaining source-unit review, then author coherent source-backed guides. Do not substitute an independently reimplemented model, a compile-only result or a simulated transport for actual C++ runtime and integration evidence.