# V2 connection-quality manuscript and native validation boundaries

This is a verified **twelve-slide isolated draft**, not a delivered fortieth
guide, narrated release, reviewed teaching assignment or completed rollback
curriculum. The full objective in `TODOS.md` remains open. The preceding goal
increment made progress through native characterization and manuscript creation;
this continuation finishes draft verification and retains its evidence. A
separate user-requested sync fast-forwarded the branch from `72c45395ac` to
`e19f0f3aa4`, preserving the unfinished presentation files byte-for-byte.

## Teaching and source boundary

The [manuscript](../../drafts/v2-network-quality.json) develops measurements,
handshake precedence, an exact two-sample example, session responsibility,
trailing pressure, fallback branch order, acceptance, policy validation,
reliability scoring and diagnostics. Its closing investigation follows an
unaccepted ranked offer through input, predicate and consumer evidence. Every
slide has explanatory notes, three reader questions and pinned sources.

Source pin: `1ceae52c7896a59ac74a4429d410f3218f0b91c6`. The entire architecture
unit `020598d821f509c36fd7427f`, lines 262–309 of
`V2/architecture/rollback-netcode-and-tag-team.md`, has been read, including its
table, diagram and related-operation claims. Its 322 words are not counted as a
reviewed teaching assignment. The complete quality analyzer and relevant
sample/feature/policy/decision/stamp definitions and predicates supply the local
implementation corrections. The learning map retains the separate session and
call-site review; no source reference is promoted to running integration.

Important distinctions include strict RTT versus inclusive loss thresholds,
maximum jitter versus average RTT/loss, trailing entries versus elapsed time,
default-policy thresholds versus configured helper thresholds, and an offered
no-contest flag versus a settled match. A stable Hybrid decision can copy an
acceptance boolean and then fail its separate accepted-state validator. A
configuration predicate can pass while the default feature policy is misaligned.
Diagnostic flags and an endpoint name do not perform a write or request.

The [authoring brief](../../authoring/v2-network-quality-guide.md) applies the
frontend skill's content-led composition: independent evidence rows, exact
comparisons, a selectable source excerpt and a genuinely ordered closing
investigation. It preserves Oshun typography, palette and existing
reduced-motion-aware interactions. There is no invented gameplay/dashboard image.

## Native execution, with the original qualifications intact

`QualityValidatorBoundaries` extends the
[nineteen-case receipt](../v2-rollback-native-cases-2026-09-07/README.md).
All **twenty** named cases are discovered and execute successfully, with zero
test errors/warnings and observed process exit zero. The report and each unique
start/success log pair are checked against the test source, not just an aggregate
count. Seven engine startup `LogFeaturePack` errors about absent preview images
remain in the raw log; they are not omitted or counted as test successes.

The new case checks default/aligned policy predicates; the accepted stable-Hybrid
counterexample; unaccepted extreme RTT; jointly offered/no-contest flags;
missing-ID diagnostic flags versus stamp validity; negative RTT with a clamped
score; and the absence of a 100-percent upper loss guard. These bounded examples
do not prove production reachability or complete input validation.

The explicitly adapted, project-local Linux fixture is unchanged except for
the host test translation unit and its generated manifest entry. All **319
copied source files**, **twelve substitutions in six files**, and the other three
host templates match the preceding receipt. A guarded one-time refresh verifies
the old Git-backed templates before replacing the test template. The action
plan admits one compile and one test-module link; both complete with observed
exit zero. Six dependency modules and three metadata records remain byte-exact;
all seven module hashes match the before-runtime list afterward. The 340 retained
compiler inputs and recorded engine identities are rechecked.

This does not repair product or engine sources, rerun the original 321-action
dependency build, make the unchanged-pin build pass, or prove Blueprint/full-game
acceptance, sockets, encryption, rendered audiovisual effects or settlement.
The original native failure and adaptation history remain in the older receipts.

## Automated verification

`state.json` binds current inputs, raw evidence, exact native identities and
the draft HTML/manuscript/PDF hashes. Distinct folders retain earlier failures;
final passing reports never overwrite failed observations.

- **216 Python tests** pass, including 22 native-preparation/integrity tests and
  seven new manuscript/PDF-checker tests. **274 Node tests** pass.
- The complete isolated draft suite passes all twelve slides at 1440×900,
  1280×720 and 390×844, with sources, transcript alignment, navigation/history,
  search/dialog/focus behavior, table-column reachability, axe and print/PDF
  checks. No runtime errors occur. Draft mode intentionally skips actual audio.
- The dedicated semantic matrix passes **60 checks**: every slide at desktop,
  compact, mobile, 320px narrow and enlarged-body mobile. Eight real DOM
  negative controls reject changed thresholds/aggregates, missing evidence,
  hidden responsibilities/annotations, changed source code, a split numeric
  heading and an unknown specimen under the canonical-only audit.
- The existing PWA source-artifact pilot passes again through the unchanged
  canonical default, including embedded states, compact active playback,
  narrow/fallback/enlarged text, print geometry and its negative controls. It
  does not establish playback or embedded integration for this new draft.
- The additional PDF checker checks all twelve pages' primary text and ordered
  visual fields, retaining inequalities, case, punctuation, numbers, units,
  evidence rows, responsibilities and annotated source. It ignores whitespace
  reflow only. Seven changed/missing-text controls and a reordered-branch
  control exercise its rejection path. Text extraction is not legibility proof.
- Ruff, scoped ESLint/Prettier, delivered/preview and draft builder freshness,
  documentation coverage freshness, repository/workflow inventory freshness,
  workflow coverage freshness and copied native-source integrity pass.

No full-workspace build was launched. Targeted checks ran serially under the
resource supervisor on the 15 GiB RAM / 31 GiB swap host, with host observations
every 25 seconds. Other workspaces' processes were left alone. Draft changes
introduce no new renderer, runtime dependency or external asset. The complete
delivered center performance/release matrix is not newly claimed here.

## Actual-size visual review and repaired defect

Author/model inspection is recorded separately from automation and willing-reader
acceptance. The final evidence set includes all twelve desktop, compact, mobile
and 1280px-wide PDF page-fit images, plus all eight mobile table-end images.
The manuscript was reviewed in narrative order. Mobile images include full-height
reading captures; they do not claim that a tall slide fits in the initial phone
viewport. Table-end captures preserve row identity beside the last value column.

The first numeric example broke “Sample” and “Aggregate” inside words at 390px
despite passing cell bounds. Final headings are **A / B / Used**, with the
caption explaining samples A/B and the value used for grading. Values, units and
notes are unchanged. A word-range guard and deliberate split-header negative
control now cover this failure. Both original and repaired mobile evidence are
retained. No text-size reduction was used.

All other desktop/mobile image bytes match the first reviewed captures; the
changed example is reinspected in its final form. Every compact slide, every PDF
page and every table-end view is inspected. The ordered table treatments serve
exact rule comparisons rather than a live transition diagram; the unnumbered
responsibilities remain peers. Dense source annotation remains readable with its
three interpretations, exact lines and full provenance. These are scoped author
judgments, not an accepted benchmark or whole-library visual pass.

Three early semantic-run failures were test-harness defects: treating the
intentionally hidden desktop scroll hint as missing content, shrinking the source
symbol in the enlargement stress case, and mutating a hidden rather than active
slide. Corrections exclude only that hint, enlarge source symbols to 24px, and
require an actual active-slide mutation before checking rejection. Later scoped
lint found undeclared browser globals and an unused destructured binding; these
were fixed and the 60/8 matrix rerun. Formatting warnings were corrected. The
first canonical regression process handle was missing after completion and no
matching process remained; a distinct final run supplies an observed exit zero.

## Continuity and next required work

All **85 delivered artifact hashes** and **409 recording hashes** match the
[contracts integration baseline](../v2-contracts-integration-2026-09-07/README.md).
The delivered catalog remains 39 guides / 409 unique slides / 435 instances /
280 cited files / 39 PDFs. Audio decoding and whole-library listening are not
rerun or claimed by a hash-continuity check.

Next: produce and inspect draft narration, run actual playback and integrated
library checks, obtain version-bound willing-reader/normal-speed listening
evidence, and only then record appropriate substantive source-unit assignments.
Continue session, transport, replication-channel, snapshot-substate and online
consumer teaching; do not treat this quality guide as complete rollback coverage.
All remaining V1–V10, whole-monorepo, generation, cross-cutting, Eve, Domains,
composition and final quality requirements in `TODOS.md` remain intact.

Reproduce the additional gates after the ordinary draft build:

```sh
python tools/presentations/build-eve-oshun.py --draft v2-network-quality
CHROME_CHANNEL=chromium OSHUN_DECK_QA_DIR=/tmp/new-quality-full node tools/presentations/tests/eve-oshun-deck.mjs --draft --draft-id v2-network-quality
CHROME_CHANNEL=chromium OSHUN_QUALITY_QA_DIR=/tmp/new-quality-semantics node tools/presentations/tests/v2-network-quality.mjs
python tools/presentations/tests/check-v2-quality-pdf.py /tmp/new-quality-full/presentation.pdf --report /tmp/new-quality-pdf.json
```

Run resource preflight and supervision as required by `AGENTS.md`; the receipt's
retention scripts are exact local-evidence collectors, not portable build setup.
