All nineteen Presentation.V2.Rollback cases execute successfully in the
explicitly adapted UE 5.5.4 diagnostic host: nineteen successes, zero test
errors or warnings, and observed process exit zero. This adds fourteen cases to
the
previous five-case run.
It is source-characterization evidence for future presentations, not a delivered
rollback guide, teaching assignment, product repair or full-game acceptance.
Inputs, build and execution#
The immutable product pin remains 1ceae52c7896a59ac74a4429d410f3218f0b91c6.
Before updating the owned fixture, all 319 copied source files and four old
templates were checked against the published manifest and commit
f12beeac5a64ac3d732993de1c3208f27722e599. Only PresentationRollbackTests.cpp
changed. The refresh script additionally ran the preparer's full extra-source
check with the verified old templates, then the current-template check after
writing the test file and input manifest.
The same twelve substitutions in six copied files remain: two reflected Blueprint-visibility annotations and ten dependency expressions. All rollback mechanism bytes remain pinned. No product or engine source is repaired. The original unchanged-pin UHT failure and later dependency failures remain in the earlier receipts; this is not an unchanged-pin or Blueprint-compatibility pass.
The first incremental action discovery exited 6 because ueagent could not
write into the newly created root-owned evidence directory. After changing only
that owned directory's owner, discovery exited zero and exported two actions:
one test translation-unit compile and one test-module link. All listed
produced/deleted paths are project-local. The generated linker script also uses
engine Intermediate scratch paths for this test module's symbols/stripped
copy; it does not rebuild an engine binary. Both discovery attempts are
retained.
The serial non-unity incremental build completes both actions with observed exit zero. It reuses six dependency modules and three module/target metadata records, whose hashes still match the previous receipt. Only the test module's hash changes. This is not a fresh execution of the earlier 321-action build.
The unattended editor then discovers and executes all nineteen exact names. Both
the raw JSON report and each started/completed log entry are checked; an
observed terminal result and the explicit exit-zero runtime log agree. All seven
module hashes remain unchanged across execution. Source integrity passes before
build and after runtime. The raw report's UTF-8 BOM and logs' original bytes are
retained. Input manifests still say compiled: false and
testsExecuted: false; result evidence lives separately in
state.json.
The runtime again reports seven missing engine template-preview images during startup. These are preserved in the full log and state, not suppressed or called an error-free engine installation. No template assets are repaired or accepted.
What the added cases actually exercise#
The original prediction, correction/live-frame, recovery, resend/ACK and budget/restore cases also rerun successfully. The fourteen additions are:
| Case suffix | Executed local boundary |
|---|---|
InnerStreamValidation |
Ten-byte header plus 29-byte sample; negative frame/cast direction and 64-bit masks round-trip; replay flag; all 39 truncated prefixes fail; trailing bytes and bounded count mismatches fail with their distinct output-array behavior. No large allocation probe. |
PacketDecodeGuards |
All 93 truncated prefixes fail and preserve caller output; bad magic/version/payload CRC and oversized payload/parity spans fail; failed encoding preserves caller bytes. |
PacketValidationLimits |
Outer trailing byte accepted; self-consistent header frames may disagree with payload; FEC index/group bounds unchecked by the predicate; selected flag constraints enforced; modified parity survives payload-only CRC. |
ParityAndDatagramRecords |
Unequal-length XOR zero-extension, duplicate cancellation and empty group; packet/datagram FEC and DTLS predicate differences; failed encoding leaves invalid datagram metadata. No encryption, send or reconstruction. |
AdaptiveInitialFramesAndGaps |
Frame-zero initial windows, missing current input, deduplication, gaps, and generic min/max/count batching versus the adaptive check. |
HandshakeThresholds |
No-sample result; exact and just-over RTT/jitter boundaries, inclusive loss boundaries; average RTT/loss versus maximum jitter; wireless warnings without automatic rank exclusion; configured thresholds. |
FallbackDecisionBranches |
Sixty trailing entries versus 59, equality interruption and non-clock timestamps; empty/pressure × ranked × accepted branches; fixed delay three; default reporting/extreme thresholds versus configured caps; stable modes. These are returned decisions, not transitions or settlement. |
ReliabilityAndDiagnosticRecords |
Empty/clean, negative and saturated term clamps, rollback threshold, averaged penalties and local stamp/export flags with a named endpoint. No persistence or request. |
SnapshotHistoryAndMissingPredecessor |
At frame 91, keys 1–91 remain and zero is evicted. Explicitly invalid short-history configurations distinguish budget-first freezing from missing-predecessor offering; a later correction-free tick can advance while retaining offer state. No production-reachability claim. |
SnapshotEstimateBoundaries |
One supplied input adds 44 byte-array bytes and 140 estimated bytes including two 48-byte records; empty exact 8192 limit and one-byte overflow; rounded-up two-fighter limits. Estimates are not heap or wire measurements. |
StructuredRestorePrecedence |
Normal local capture restores the original state hash; prior presentation events remain; contradictory input uses the last matching attribute record, ignores an unknown fighter ID and refreshes fighter hashes despite inconsistent bytes/hash. Not every sub-state. |
PresentationInterpolationClock |
Fixed 1/60 step independent of refresh metadata or a four-frame gap; X/Y halfway while health uses target; negative delta, changed refresh metadata, non-increasing frames and combined fighter/player matching. No rendering. |
PresentationEventConsumption |
Non-Emit record flags with unchanged tested simulation hashes; duplicate event submission; first-tick consumption versus retained last-frame data; clamped recovery-overlay metadata. No sound/particle/UI consumer. |
CorrectionPresentationPolicy |
Three original emissions plus a live frame; two replay records emitted or suppressed according to config; correction telemetry consumed once. |
The pinned Source/Plugins exact-name caller search and inspected wrapper bodies
are source evidence only. BuildPresentFrameFromSnapshots creates a fresh local
presentation world, submits two snapshots without events, ticks once and returns
a value; it is not a display-loop consumer of the session's private world. The
search does not cover Blueprint assets, dynamic dispatch or external consumers.
Verification and remaining scope#
All 209 presentation Python tests (including 22 preparation/integrity tests) and 274 Node tests pass. Ruff, scoped formatting, builder/preview freshness, document coverage freshness, repository inventory freshness, workflow inventory freshness and workflow-coverage freshness pass. Checks run serially with resource supervision; no broad Nx, engine rebuild or production-build sweep runs.
All 85 delivered artifact and 409 recording hashes from the contracts integration receipt remain unchanged. No browser, PDF export, audio decode, listening or reader review is newly claimed. No UI, narration or teaching assignment changes.
Exact manifests, previous-receipt identity, compiler/link inputs, module hashes,
both discovery logs, build/runtime logs and raw automation report are bound by
state.json. The retained verification script checks local receipt/input bytes,
the staged versions when requested, and the still-present temporary fixture.
The complete V2 target, strict-FP/cross-platform determinism, authored assets, packed-key limits, every snapshot sub-state, channels/replication, tag mechanics, socket delivery, authenticated DTLS, FEC reconstruction, live display consumers, settlement and release acceptance are not proved. Continue source-unit review, coherent teaching and all remaining TODOs; the full objective remains open.