# Shared architecture narration and source review — 2026-09-06

The [full checklist](../../TODOS.md) remains active. This is a partial narration
and teaching increment, not complete whole-monorepo coverage, live Eve
integration, Domains acceptance or the final visual/listening review. Source
claims retain revision `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

## Current audio

Four sequential one-thread CPU batches rendered 20 previously pending tracks,
including the corrected partial-failure, event-bus and curation explanations.
All four processes exited normally with code 0. The stock Kokoro voice, model
revision, pronunciation and timing policy are unchanged:
[architecture 1](architecture-narration-1.txt),
[architecture 2](architecture-narration-2.txt),
[communication 1](communication-narration-1.txt),
[communication 2](communication-narration-2.txt).

The [current-audio audit](current-audio-check.json) verifies **145 tracks,
11,898.803 seconds and 1,845 sentence segments** against actual speech
fingerprints, hashes/bytes, decoded duration and encoded pauses. **236 tracks
remain pending: 216 stale and 20 missing.** The architecture and
communication/discovery guides now have complete audio. These mechanical checks
do not establish a normal-speed listening review.

## Standalone playback verification

Both non-draft standalone suites exited normally with code 0 and passed all **21
slide instances**:

| Guide                         | Slides | Report                                         |
| ----------------------------- | -----: | ---------------------------------------------- |
| Contracts, ownership and data |     11 | [Result](architecture-browser.json)            |
| Events, jobs and discovery    |     10 | [Result](communication-discovery-browser.json) |

Adjacent `-browser.txt` logs retain the runs. Coverage includes
desktop/compact/mobile layout, axe accessibility, source/script alignment,
navigation, dialogs, print sequence, actual decode/play/pause/resume/seek/speed
and end transitions, HTTP failures, missing/corrupt audio, retry, playback races
and portable local-file behavior. No narrator was running during these checks.
The reports bind the exact delivered HTML, not a subsequently rebuilt variant.

The [complete center run](center-ready-audio.json) also exited normally with
code 0. It covers **38 guides, 407 embedded slide instances and 38 PDF
exports**, with 814 desktop/mobile initial-viewport captures. All capture bytes,
45 bound center artifacts and the 38 delivered PDFs were revalidated against the
completed report. The run uses `--draft --ready-audio`: 30 desktop/mobile
playback boundaries cover the 15 ready guides; the remaining 23 guides are
explicitly excluded from audio acceptance. Navigation, hierarchy, listening-time
semantics, search, browser history, drawer/focus handling and the local
interaction budget pass. This is not full-library audio acceptance.

## Targeted visual review

The [visual ledger](visual-review.json) binds **25 inspected images across four
unique slides and four guide contexts** to their canonical slide, HTML, PDF and
narration hashes. Partial failure, its closing recovery table, event-bus
mechanisms and curation were inspected in desktop, compact desktop, full-height
mobile, embedded desktop/mobile initial viewports and actual final PDF-page
renders. The recovery table also has a mobile table-end view. The mobile
embedded image is only the initial viewport; it does not establish a full-scroll
review on its own. Notes and questions were reviewed for all four slides.

No material defect was found in these inspected states. The final all-slide
visual and normal-speed listening reviews remain open. Repository Playwright,
axe and Poppler supplied the available browser, accessibility and PDF workflow;
no dedicated presentation/PDF skill or audio-listening tool was available.

## Source fidelity

The [learning map](../../drafts/shared-architecture-learning-map.md) records the
source review and remaining depth. Two new reviewed assignments cover complete
source units, without automatically closing neighboring units:

- [The partial-failure contract](../../assignments/architecture-partial-failure.json)
  permits empty results with errors and `partial: true`. The flag means error
  presence, not a successful effect. Corrected explanations distinguish schema
  validation, generic result/error schemas, builder inference and the weaker
  structural guard. The closing recovery table no longer invents success.
- [The curation package boundary](../../assignments/shared-curation-surface.json)
  identifies the private 0.1.0 package, absent declared runtime dependencies and
  its module responsibilities. The code has six collection kinds and nine
  exported module groups, not the seven kinds and six groups stated in the
  architecture page. Review, importers and web clipper are the omitted groups.
  Generated artifacts and HTML/oEmbed indicators do not establish persisted
  sharing, transport, erasure, audit or actual external embed preservation.

The event-bus explanation also corrects recovery implications. Publication IDs
and replay counts do not establish consumer completion. Replay can be disabled;
retention and group claims use the configured event lifetime. The pinned
scheduler performs separate read/remove/publish commands despite a comment
describing a Lua transaction. The tests execute original methods with controlled
Redis replies and a manually triggered callback; they do not exercise a Redis
server, real timer, restart, network or deployed worker. No whole event-bus unit
is newly assigned by that limited explanation.

The frontend skill guided retention of the fan-out diagram, plain event-bus
columns, the recovery table and the keep-to-revisit flow. The fan-out picture is
one mixed-result example, not a universal successful-result guarantee. Exact
package inventories and code counterexamples remain in notes and source context,
without crowding the canvas or adding decorative motion. No app/service runtime
behavior or live provider acceptance is claimed by these presentation changes.

## Automated checks and reproduction

All **54 Python tests** and **37 Node tests** pass in
[the Python log](unit-tests.txt) and [the Node log](node-tests.txt). Fifteen new
Node regressions cover seven pinned partial-failure schema/helper cases, five
mocked event-bus command/return boundaries and three curation inventory and
authored-explanation checks. The partial-failure tests use the contracts
package's installed Zod 4.4.3; the source pin declares a compatible 4.4 range.

The explicit [teaching-completeness gate](teaching-incomplete.txt) exits 1 as
expected: 11 reviewed units, 103,315 unassigned meaningful units and no invalid
or stale assignment errors. It is incomplete, not a passing global release gate.

Builder and teaching-inventory freshness, scoped Ruff/ESLint/Prettier and
whitespace checks pass in [the static log](static-checks.txt). The
[receipt](checks.json) binds current artifacts, reports, logs, visual review and
relevant tools, with all global completion claims false. Historical receipts
remain historical checkpoints, not evidence for subsequently changed bytes.

Use the configured Python environment and browser before reproducing:

```bash
python -m unittest discover -s tools/presentations/tests -p 'test_*.py'
node --test tools/presentations/tests/*.test.mjs
python tools/presentations/build-eve-oshun.py --check --preview
python tools/presentations/check-center-coverage.py --check
python tools/presentations/check-center-coverage.py --require-complete
```

The final command is intentionally incomplete at this checkpoint. Run browser
suites sequentially with the repository supervisor and no narrator, using
`node tools/presentations/tests/eve-oshun-deck.mjs --guide architecture`, then
`--guide communication-discovery`; run
`node tools/presentations/tests/presentation-center.mjs --draft --ready-audio`
afterward. Export all center PDFs last so the report binds the delivered files.
Use separate temporary QA directories, `CHROME_CHANNEL=chromium` on this host,
and the resource checks required by `AGENTS.md`. The observed browser was
Chromium 148.0.7778.96 with Playwright 1.60.0 and axe 4.11.4.

For the current-audio subset, pass the audit's comma-separated `currentIds` to
`render-eve-oshun-narration.py --check --ids`. The unfiltered full-inventory
narration gate remains required after all pending tracks are finished.
