# Agentic source application and current-library automated checkpoint

Source pin: `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.
Application baseline: `7167e08ce5cb61b11995f1541a18ed960e6e61d1`.

Eight corrections and three new explanations are applied, and the complete
current-library automated checks pass. This is not whole-guide/source
acceptance, normal-speed listening, reader evidence or completion of the TODO
objective. The V2 mobile label finding below remains unresolved.

## Applied explanation changes

The [application record](../../authoring/agentic-source-application.json) binds
eight replacements and three new slides. Recovery distinguishes selected-field
matching from full content integrity, budget admission and durable dispatch.
Preflight, grant selection and handoff distinguish supplied declarations and
local helper checks from authenticated authority and integrated enforcement.
Admission compares the helper, legacy loop and governed route. The new eligible
execution path, outcome accounting and interrupted completion example separate
an attempted call, a saved record and an external effect.

The frontend skill informed the bounded visual changes: independent envelope,
handoff and interruption comparisons omit ordinal markers; the true five-step
execution flow retains them. Removing the obsolete admission SVG override lets
the authored responsibility table appear. Other authored slides and their order
are preserved. This is not acceptance of the complete Agentic guide, the
composition pilot or the library-wide redesign.

Historical proposal tests keep the immutable preapplication baseline. Separate
live-content tests require the actual eight-plus-three application, only the
three declared numbering refinements and one retired diagram override, with all
380 other authored slides unchanged. Old proposal text is not rewritten to match
an accidental implementation. No new section-level teaching assignment is
claimed from the new citations alone.

## Verified narration and standalone bank

The completed 388-track queue was followed by eight corrected and three new
recordings. The [final renderer log](logs/narration-applied.log) records the full
391-track fingerprint, byte/hash, duration, decode and encoded-pause check:
34,841.439 unique seconds and 5,167 segments. Following all guides totals
36,918.761 seconds because shared context plays again in each guide.

[Continuity](audio-continuity.json) separately verifies identical metadata and
bytes for all 380 unaffected recordings, including all original 177 current
tracks. None of the eight replacements belonged to that original current
subset; the contrary historical recovery exception was incorrect. Current
review prose is corrected, without rewriting old receipt evidence.

The [full standalone report](standalone-bank.json) completed with no errors:
391 slides at desktop, compact and mobile sizes; source/transcript alignment;
accessibility, navigation, dialogs and print/PDF sequence; actual playback,
pause/resume, seek, speed and end transitions for every track; HTTP failure and
race scenarios; and stable tested HTML. Forty-four additional observations
cover all eleven applied slides in four states, including compact playback.

One earlier bank launch overlapped the final builder by three seconds. Its
owned process group was stopped and returned 143; its partial log is retained
as `logs/bank-before-stable.log`, not passing evidence. The completed bank run
started after the builder terminated and its freshness check passed.

## Embedded center, focused guides and PDF checks

The complete, non-draft [center suite](presentation-center.json) passes all 38
guides and 417 guide-slide instances in both viewports: 834 verified initial
captures, 76 guide playback/end/replay boundaries, 38 PDF exports, five complete
hierarchy-duration audits and 38 selected-guide/detail duration checks. It also
passes search, navigation/history, collapse persistence, mobile drawer/focus,
guide-switch playback cancellation, accessibility, performance budgets and
tested-artifact stability. Twenty-two extra observations cover the applied
Agentic explanations in both embedded viewports.

The complete non-draft [V2 promise](v2-product-promise.json) and
[V2 surfaces](v2-product-surfaces.json) standalone suites pass nine and twelve
slides respectively, including actual narration. The
[Agentic PDF content gate](logs/agentic-pdf-content.log) passes all eleven
affected pages: visible characters, exact comparison cells, flow order and the
absence of ordinal markers on independent comparisons. Its five unit cases
include deliberate text/order/numbering failures. The preliminary PDF correctly
failed on the envelope's old ordinal markers; see `logs/agentic-pdf-before.log`.

The final presentation unit suites pass 179 Python and 268 Node tests. The
Agentic package's 360 tests, 104 pinned-source/configuration probes and 25 BFF
route tests also pass; no application implementation or typecheck baseline was
changed. Ruff, scoped ESLint/Prettier, builder and repository/workflow/assignment
report freshness pass. Full teaching gates still deliberately fail: 103,300
meaningful documentation units and 3,718 workflow units remain unassigned.
Twenty-six documentation units are reviewed; no source-family backlog is hidden
behind the current narrated-slide count.

## Scoped visual review and unresolved findings

[The model review ledger](visual-review.json) retains 93 named captures: 23
Agentic standalone states, all 17 Agentic PDF pages, all 21 V2 PDF pages,
22 affected Agentic embedded states, eight representative V2 embedded states
and two V2 full-height reading states. Sixteen Agentic PDF pages are byte-identical
to the pages inspected earlier in this phase; the
[continuity comparison](pdf-review-continuity.json) records that exact match.
The final envelope page was inspected again after removing its ordinal markers.
No full-library model review or human acceptance is claimed.

The V2 companion table's `Authentication contract` row heading breaks mid-word
at phone width (`Authenticatio` / `n contract`). Its fields remain present and
the automated bounds checks pass, but that does not make the wrap acceptable.
The immediate follow-up is to shorten the visible label without reducing font
size, retain the full OAuth explanation, review affected teaching fingerprints
and repeat invalidated browser/mobile/PDF and narration-freshness checks before
integrating the saved contracts guide. Keep this finding open in TODOs.

The existing V2 promise audience comparison and several surface comparisons
retain legacy ordinal markers. The embedded ownership diagram also has smaller
labels than surrounding prose. Their full content-led composition/typography
and willing-reader review remain open; this delivery checkpoint does not promote
them into an accepted benchmark.

Named visual captures are model inspection only. Tall mobile reading/table-end
images are not proof that the entire slide fits inside a phone's initial
viewport. Automated PDF text extraction checks character presence and order,
not optical word spacing or legibility. Neither screenshots nor playback
automation establish normal-speed listening or willing-reader comprehension.
Those reviews remain open, as do exhaustive source teaching, production
durability, the full composition/craft standard and the broader TODO objective.

## Reproduction and receipt integrity

Follow repository resource checks and use the prepared Python dependencies.
Run expensive gates serially through `tools/presentations/run-supervised.py`.
The complete browser commands have no draft, smoke, audio-only or layout-only
flag; use distinct QA output directories to preserve each scope.

`collect.py WORKDIR` only retains already-produced reports, logs, selected
images and named input hashes. It does not run tests or certify acceptance.
Its generated [state](state.json) binds the named inputs and retained evidence.
Run `verify.py` and `test_verify.py`. Their checks reject incomplete scopes,
missing/reordered results, duplicated coverage, missing guide audio, changed
HTML/PDF bytes and changed reviewed images. They do not infer human findings
from an automated pass. The ignored whole-bank fixture can be rebuilt with
`build-eve-oshun.py --preview`; its hash is checked when present and it is not a
public monolithic reading route.

The previous canonical browser reports are preserved in
`historical-current-reports/`; their canonical paths now contain this phase's
complete reports. Historical `summary.json` is unchanged and remains historical.
Run final receipt checks again after normal commit hooks if any bound file changes.
