# V2 rollback-session integration · 2026-09-07

This receipt records the fourteen-slide **V2: correcting a predicted frame**
guide, its recordings and PDF, two complete architecture teaching assignments,
and five added native characterization cases. It does not close the exhaustive
presentation objective or establish reader/listening acceptance. The
[state](state.json) binds the final inputs, evidence and observed command exits.

The guide belongs under Products → V2 competitive fighting → Netcode, with
product promise and contracts as prerequisites. Its
[brief](../../authoring/v2-rollback-session.md) gives every slide a
communication job and keeps all eight craft requirements explicit. One worked
tick carries the explanation: current frame 4, corrections for 3 and 2, restore
1, replay 2–4, then advance and return live frame 5. Its correction-end hash and
returned live hash deliberately describe different frames.

## Source and native evidence

The source pin is `1ceae52c7896a59ac74a4429d410f3218f0b91c6`. The complete
session and predictor implementations and their headers were reviewed, together
with the exact equality/hash, correction, configuration and telemetry
predicates. The two assigned architecture sections span lines 89–147 and 148–168
of `V2/architecture/rollback-netcode-and-tag-team.md`. Their exact unit hashes
and explanatory slide fingerprints are in the
[assignments](../../assignments/v2-rollback-session.json). Neighboring sections,
generated duplicates and C++ semantic units are not assigned by this work.

The guide distinguishes a stored prediction from an already consumed comparison
point, the final confirmed payload from the mismatch collected earlier, and a
successful replay from a frozen retry. It teaches the recomputed 32-bit hash's
specific scope, packed sixteen-bit player identities, unused stored prediction
horizon, caller validation boundaries and copied net-graph measurements. Invalid
player indices are deliberately supplied through local calls; the examples do
not establish an externally reachable packet path or live-match exploit.

The [native report](native/report.json) records 25 successful tests, zero test
warnings/errors, zero failed/not-run/in-process cases and observed process exit
zero. `ConfirmationLifecycle`, `PackedPlayerKeyBoundary`,
`EarliestCorrectionAndLiveHash`, `DuplicateConfirmationOrdering` and
`FrozenConfirmationRetryAndReset` are new. Replay and live hashes are compared
against calls to the actual C++ simulator, including both correction orders.

The existing explicitly adapted Linux UE 5.5.4 fixture retains all 319 pinned
source files, four templates and twelve substitutions in six dependency files.
Only the test template changes. The exported action plan contains one compile
and one link; listed outputs/deletions stay in the owned temporary project. Its
link script still uses engine Intermediate scratch for the test module. All six
dependency modules and three metadata records remain byte-identical, and all
previously recorded engine/compiler inputs were checked again. No product-source
repair or engine binary rebuild is included.

The [previous draft receipt](../v2-network-quality-draft-2026-09-07/README.md)
retains the build lineage and original unchanged-pin UHT/dependency failures.
Seven missing engine feature-pack preview images still produce startup errors;
they remain visible in the retained engine log, separate from the successful
test entries. This is not an unchanged-pin/full-game build, Blueprint
acceptance, strict-FP/cross-platform determinism, real socket/authentication
acceptance or an actual sound, particles or HUD consumer test. Preparation's
`compiled: false` and `testsExecuted: false` flags remain preparation semantics;
execution results are recorded separately.

## Automated delivery and preservation

The required observed results in [state](state.json) cover the full 484-slide
bank, all 48 center guides and PDFs, the complete recorded standalone guide,
every new slide in the focused semantic/reflow/playback matrix, all 138 tables
and actual browser zoom, and the catalog-wide short-viewport regression. Full
center verification includes all 507 guide-slide instances and 96 actual-audio
boundaries. These are complete non-draft runs, with no subset flags replacing
the bank or center gate.

The focused suite checks exact authored fields, source characters and annotation
order, visible content and overflow, 320px/390px/landscape/enlarged-text
contexts, expanded/collapsed embedded workspaces and actual recorded playback.
Every new table is also selected for embedded 200%/400% browser-zoom probes.
Deliberate defects exercise changed frame values, missing
source/rows/annotations, hidden controls, reduced headers and missing title ink.
The zoom extension checks the actual browser setting, DPR and physical
screenshot geometry; CSS enlargement is recorded separately.

Python and serial Node suites, Ruff, scoped ESLint/Prettier, source-unit and
inventory freshness, builder/preview/draft freshness, exact PDF content and all
484 audio fingerprints/hashes/durations/decodes/encoded pauses are required.
Presentation accessibility and local interaction timing are included in the
browser gates. No unrelated Nx/mobile/production-app build matrix is implied.
The documentation and workflow `--require-complete` gates deliberately exit one:
40 documentation units are reviewed, 10,018 structural and **103,286
unassigned**; all **3,718 workflow units remain unassigned**. Freshness passing
is not complete teaching coverage.

[Preservation](preservation.json) checks all 470 prior complete audio entries,
recording hashes and speech records, all previous canonical content files and
all 47 prior guide definitions and standalone HTML bytes against
`8c27a8d3350fe7484c18585b570f72a561090540`. Exactly three newly cited paths
change repository citation flags; pinned file records and workflow declaration
units remain unchanged. The fourteen new recordings add 1,237.964 seconds. The
complete library totals 43,762.106 unique audio seconds and 45,640.455 route
seconds, since shared slides play again on their guide routes.

All 48 PDFs were exported by the full center run. Its regenerated prior PDFs
changed beyond the permitted date and structure-node normalization, so the 47
prior deliverables were restored byte-for-byte from the baseline. The final
[prior-PDF comparison](pdf-preservation.json) requires their exact hashes. The
new guide's extracted PDF text preserves every non-whitespace character and
visual-field order; this is separate from optical review.

## Review, failed attempts and limits

The initial PDF gate rejected a valid line wrap after a C++ cast. The shared
source-page checker now normalizes extraction whitespace while retaining all
non-whitespace characters, provenance and annotation order. Regression cases
reject changed casts, missing operators and reordered annotations. Exact
selectable HTML source still includes the original whitespace. The initial
failure and corrected draft run remain in history.

The first integrated control check scrolled to the Pause button and incorrectly
required the seek control below it to be onscreen. The complete normal-flow
footer is reachable by scrolling; the corrected check brings that region into
view and tests every control. Its failed exit and before/after images remain
retained. Some development tool versions were not frozen separately; those
historical runs do not replace the final checks bound to current tool inputs.

Optical inspection then exposed compact screenshot pixels at an older scroll
position despite current DOM coordinates. The new focused capture path requires
consecutive stable static bitmaps and title ink at the declared coordinates;
playing captures warm the compositor and record their dynamic scope. This is a
capture correction, with no shared CSS/JS or product rendering change. The
earlier compact images are development evidence, not final viewport-fit proof.
Full-bank/center layout assertions remain useful; their uninspected images are
not promoted to a new whole-library optical review.

Two later semantic browser children completed their full 238-check, 254-capture
run, but the supervising shell was terminated before it could retain an observed
exit record. A detached retry was terminated at startup. Their complete reports
and available logs remain in history as interrupted-run evidence only. The final
supervised `semantic-stable` rerun completed with observed exit zero and is the
sole semantic result used by the required gate and visual ledger.

During the final integration, `origin/main` advanced by seven Eve guides and
their generated catalog routes. The rollback work was rebased onto that exact
baseline, then every affected catalog, source, audio, browser and PDF check was
rerun. Two Python attempts exposed stale-baseline and supplemental-source
assumptions; the rollback preservation test now assembles every baseline source
form. Builder and workflow freshness checks also correctly rejected generated
files until their canonical write passes completed. Failed attempts and their
observed exits remain in the receipt; only the successful final names are
required by [state](state.json).

The first post-merge PDF preservation run rejected regenerated bytes for prior
guides. A failing `portfolio-map.pdf` specimen is retained in history. Restoring
all 47 baseline PDFs exactly produced a successful comparison while the new
rollback PDF independently passed text extraction and page-by-page review.

Four visual labels were corrected during editorial inspection: the caller does
not supply the completed next-frame input set; mixed correction/restore values
need a Record/Value/Role heading; hash labels can be shorter; and the net-graph
arithmetic must identify players 0 and 1. Every spoken block is preserved across
these edits. The draft JSON retains its historical authoring-status sentence;
canonical delivery and this receipt establish the integrated state.

The [visual ledger](visual-review.json) names the images actually inspected and
their exact bindings. Of its 186 images, 140 are exact-pixel matches to the
pre-merge review and 46 changed embedded captures were inspected again after the
merge. All fourteen PDF pages were reviewed at 960px page-fit width. Source
identifiers can wrap; explicit line numbers and annotations retain their
relationship. Tall mobile reading images and intentional footer/table scrolls
are not claims that all content fits the first viewport. Automation, optical
model inspection, normal-speed listening and willing-reader comprehension remain
distinct forms of evidence. The latter two, an accepted visual benchmark and
complete snapshot/combat/transport/consumer coverage remain open in
[TODOS](../../TODOS.md).

Resource checks and supervised runs are serial for expensive targets. Retained
logs remove PID-prefixed process-list rows, ANSI and trailing whitespace, with
raw/retained hashes and transformation counts recorded. The runtime supervisor
recorded the editor's exit before its trace-server descendant appended shutdown
diagnostics to inherited stdout; both the at-exit prefix and post-exit append
are hashed separately. The runtime index keeps its original bytes; the separate
engine text normalizes line endings and trailing whitespace. No secrets file was
accessed for this task.

`scripts/verify-retained.py` checks these retained hashes and required observed
results; it does not rerun native/browser tests or establish new acceptance.
