# Offline continuity — reviewed source findings, 8 September 2026

This is the twenty-slide guide’s source and delivery review. Four complete
PWA/offline documentation units have reviewed teaching assignments. Primary pin:
1ceae52c7896a59ac74a4429d410f3218f0b91c6. The cited implementation files are
byte-identical at authoring baseline 85f3a7aca42b4c773d5b5e96d38b201c8b9b1dc4.
This is source continuity, not deployed behavior certification.

Read in full: web public/sw.js; pwa-service-worker.ts; pwa-runtime.ts;
pwa-recent-content.ts; offline-shell-routes.ts; PwaBootstrap.tsx; manifest.json;
offline library types, queue, cache, retry, storage, connectivity and index;
`offline/src/__tests__/offline.test.ts`; pwa-service-worker.test.ts. The
complete policy test was also reviewed, including classifier, marker and
activation cases. Product Surfaces feature PWA/offline sections and
high-level-architecture offline section read; entire source files are not yet
assigned as fully reviewed.

## Distinct implementations

Product Surfaces describes queueBackgroundAction and mobile behavior as leaning
on @oshun/offline. The pinned web worker implements its own IndexedDB queue and
fetch replay. It does not import OfflineSyncQueue. Its delete/retry rules differ
materially from the reusable library. Show both implementations separately and
trace actual callers before saying a surface inherits the library's semantics.

The web helper posts OSHUN_QUEUE_ACTION and returns true; no message-port ack or
await of IndexedDB persistence exists. The worker persists pending-actions with
auto-increment local IDs, then broadcasts pending count and registers background
sync if supported. Non-GET fetches are not automatically queued. No alternative
replay trigger is implemented in this worker when Background Sync is absent.

Web replay reads before deleting and fetches with credentials include. Network
throw stops the drain and preserves current/following actions. 2xx deletes and
increments replayed. Every 4xx deletes WITHOUT incrementing replayed, including
401/403/429; there is no dead-letter record in this implementation. 5xx retains
that action but continues to later actions; no per-object ordering fence,
exponential backoff or retry ceiling is applied here. A network response
followed by failed local deletion can repeat delivery; the caller/server's
idempotency contract must settle duplicates. Pending zero alone is not proof all
writes succeeded. The generic SW status contains only replayed/pending counts.

## Install, update and restore

Manifest is Lilith, four shortcuts (Tara, Nyx, Arete, Nisaba), standalone start
/?surface=pwa; two screenshot labels retain stale six-room wording. Release
admission follows brand/registry consumers, not labels.

Registration guard retries after failure, handles post-hydration registration,
and exposes waiting workers only when another controller already controls the
page. sw.js install calls skipWaiting immediately and precaches each asset
best-effort. Activation deletes old oshun-\* buckets, claims clients and
broadcasts OSHUN_SW_UPDATE. The renderer must not teach that Keep reading
prevents worker activation. PwaBootstrap's controllerchange handler avoids page
reload without local or sibling-tab apply intent, and turns the prompt ready
instead.

The prompt is deferred until cookie consent has a decision. Dismissal suppresses
that update token. A waiting-phase apply saves route/scroll, sets and broadcasts
intent BEFORE postMessage, then uses controllerchange or a four-second reload
fallback. A ready-phase apply reloads directly. Sibling-tab intent can cause
this tab to reload at controllerchange; sibling scroll is not saved by that
handler.

Scroll restore stores path, timestamp and positive rounded Y in sessionStorage.
It accepts the same path within 60 seconds, tolerates an initial missing query
without deleting the snapshot, and retries every 120 ms up to six seconds until
within 16 px (clamped to available scroll height). The first restore clears the
persisted snapshot; later retries use the closure. This is scroll restoration,
not recovery of unsaved in-memory edits, media position or an acknowledged
write.

Standalone launch normalizes/sanitizes targets, excludes public shell paths and
removes surface=pwa. LocalStorage plus a 30-day SameSite=Lax cookie retain the
route; updatedAt is stored but not used for expiry validation in this helper.
Loading/auth and explicit current location matter: /profile is preserved,
specific deep links are preserved, signed-in root/public entry may restore the
saved target, signed-out entry can route through welcome?redirect. An existing
welcome redirect is respected. Storage operations themselves can still throw
where the write method is not caught; do not promise failure-proof persistence.

## What is cached

Fetch classifier precedence: non-GET/non-http ignored; auth/session
network-only; Nisaba passage-daily/notebooks path prefixes allowlisted; every
other API-shaped request passes through network; media cache-first; warmable
HTML and shell navigation network-first; other .js/.css/.json or \_next assets
stale-while- revalidate. This is a code policy, not a claim every deployed
backend is safe or every endpoint current. Prefix matching includes descendants.

Nisaba cached reads return immediately while network revalidates, except that a
Date header older than 24 hours makes them await network. No network response
then gives 503 NISABA_OFFLINE_CACHE_EXPIRED; no cached item/network gives 503
NISABA_OFFLINE_CACHE_MISS. A missing/invalid Date does not trip the age branch;
at exactly 24 hours the comparison is not expired. Network error HTTP responses
are returned as responses; only thrown fetch/cache-chain errors become null. Do
not teach a guaranteed 24-hour maximum without that qualification.

Shell navigation uses successful online response, then on thrown failure exact
cached request (ignoreVary), then cached home with requested-URL marker and
static recovery controls, then minimal offline HTML with status 200. Non-2xx
HTTP responses return directly instead of entering this fallback. Exact-cache
returns do not have the home-fallback marker injected. Home substitution does
not establish that the requested object is available.

Media has a 200-entry cap; BFF allowlist 50; other runtime assets 100. Cache
keys are deleted in cache.keys order; there is no read-time touch implementing
true LRU. Static shell cache has no matching entry cap in this code. No
fabricated cache-hit, storage-byte or performance measurements should appear on
slides.

Recent content stores up to eight normalized path records, with title/summary,
source and cachedAt; bootstrap records a route BEFORE warm success is known.
Fallback merges those with Nisaba recent passages, sorts and dedupes by path; up
to four recent routes contribute to warming alongside four core routes and the
current eligible route. The warm helper returns a successful non-redirect HTTP
count, not cache inspection. Bootstrap's ready flag means count > 0, not that
every route/asset or recent record is recoverable. Repeating the same route
signature reuses the previous warm result.

## Shared queue and storage limits

OfflineCache JSON-stores value, updatedAt and optional expiresAt. GET deletes
expired entries; stats counts them without deleting. Missing TTL means no
expiry. MemoryOfflineStorage is a Map; cross-process persistence requires
another adapter. Cache/queue constructors accept storage, clock and (queue) ID
seams.

OfflineSyncQueue stores whole active array at one key, dead letters at another.
No transaction/lock wraps concurrent enqueue/process or cross-key moves. A
fulfilled handler removes an item; caller decides what counts as fulfillment.
Default retry is four total failed attempts with 1 s, 2 s, 4 s scheduled delays;
formula grows by 2 up to 30 s under longer custom policies. Unclassified errors
retry; caller classifier may send deterministic failures directly to dead
letter. Dead letters retain whole item, message, attempts, reason and timestamp.

Within a process() pass, deferred/retry/dead items block later equal objectKeys.
Keyless work is independent. A retry stays at its prior queue position and
continues blocking next passes until due. Dead letters are not consulted to
rebuild blockedKeys in a later pass. replayDeadLetters appends to active queue,
resets attempts/due time, retains original id and createdAt, then removes the
selected dead letters; it does not restore prior relative order. Reusing an ID
only enables server idempotency when the actual transport sends and the server
enforces it. Do not state global ordering or atomic exactly-once durability.

process saves the next active queue BEFORE saving new dead letters. Replay saves
the active queue BEFORE removing dead letters. Failure/crash between those
writes can lose a newly failed record or leave replayed copies in both stores.
The tested happy persistence path remains useful; it is not a transaction or
crash-atomic proof. canProcessQueue is a separate caller utility: online
required, metered denied unless allowMetered true; process itself does not
invoke it.

## Member-data clearing

Client marks that a member session may have cached bytes. On member-to-anonymous
clear it posts OSHUN_CLEAR_MEMBER_DATA if a controller and marker exist, then
removes the marker before any worker acknowledgement. Inaccessible localStorage
reads assume a purge is needed. Pure anonymous transitions without a marker skip
the purge to preserve public cached shell data.

Worker uses allSettled to delete BFF, static and media buckets and queued
actions, then re-precaches public install routes. Runtime assets are retained.
Individual purge failures are swallowed by allSettled and no result ack returns.
This does not clear every localStorage/Iris/third-party cache; each owner needs
its own cleanup. Cached authenticated HTML and media still demand shared-device
review.

## Authoring package · verified for stated technical scope

Purpose: teach product, design and engineering readers how a member returns to
reading, understands cached evidence, and follows a queued action to its owner.
Prerequisite: `v1-web-pwa`. Home: Products → Oshun V1 → V1.0 → Member
experience, after account controls. Guide ID: `member-offline-continuity`. The
twenty-slide manuscript is in `content/30-offline-continuity.json`. The final
narrated HTML and PDF have passed the affected technical review. The full V1
objective remains active in [v1-completion.md](v1-completion.md).

Visual thesis: a calm paper field guide with readable native diagrams, explicit
branch labels and terracotta relationships. Compare equal outcomes as peers;
never highlight an ordinal row by default. Field specimens contain actual names
without invented member records. Dense policies receive their own readable
table.

Content plan: open with a reading return; explain distinct runtime owners,
installation and update intent, route restoration, fetch precedence, freshness,
document fallback, recent warming and cache limits; then follow web submission
and replay before explaining library storage, records, retries, ordering, replay
and crash boundaries. Close with session clearing and the original reader task.

Interaction thesis: preserve source disclosure, reduced-motion-aware navigation
and workspace transitions. Diagrams and complete explanation remain usable in
static PDF and mobile reading order. No shared renderer, CSS or player changes
are planned.

Deliverables: canonical content and catalog membership after candidate review;
source-unit assignments only for complete reviewed units; focused review config,
HTML, twenty new recordings, PDF, catalog duration updates and a retained
receipt. Preserve all 600 upstream tracks and all 61 pre-existing guide HTML
files against the immutable baseline. Scope supporting code claims separately
from full V1 source coverage. The four feature/architecture PWA and
offline-substrate units have reviewed assignments, with their old manifest,
common-queue and `flush` claims explicitly reconciled against code.

Acceptance: each slide explains one useful mechanism and has three developed
paragraphs and three review questions. Inspect all rendered slides; keep age,
rejection, missing acknowledgement and storage failure qualifications visible.
Use pinned-source characterization fixtures for replay/status, expiry and queue
ordering/crash boundaries. These fixtures do not certify a deployed device.

Verification: build the draft and run
`eve-oshun-deck.mjs --draft --draft-id v1-offline-continuity --layout-only`;
inspect actual desktop/mobile captures. Then run the configured guide review
with four standalone and three embedded views, enlarged text, axe and
deliberately removed evidence. Run center review with
`--guides member-offline-continuity --base-revision 949d68e86ef6477118d446e7f5c9d2e0b09aa4dc`,
the source-map PDF checker, relevant presentation unit tests,
lint/format/build/coverage freshness and encoded narration verification. Export
and inspect the exact final PDF. Use Python 3.11, one CPU narration worker and
supervised serial expensive checks on this host.

Reader tasks: distinguish worker activation from page reload; interpret an
unknown cache age; explain pending zero after rejection; follow same-object
retry ordering and dead-letter replay; identify the missing clear
acknowledgement. No human reader, normal-speed listening or new live-device
session has yet been performed. Do not count automated or model optical review
as that acceptance.

Recovery and publication: isolate work on the existing V1 task branch, stop only
owned jobs, preserve valid partial audio and retain failures. Keep unpublished
content in draft until candidate review; never advertise missing deliverables as
ready. Commit with normal hooks and push to origin/main after applicable
technical checks pass. Publication of this increment does not close the complete
V1 task.

## Candidate findings and corrections

The twenty-slide desktop sequence and native-size slides 3, 7, 11 and 19 were
inspected by the model after the draft browser/PDF pass. The web-enqueue return
arrow originally combined pending-status broadcast with sync registration even
though the latter goes to the browser, not the calling feature. Its final label
now names only the pending-count broadcast; narration explains registration
separately. Owner descriptions were shortened to preserve spacing between
columns. The three ownership accents follow worker activation, stored action and
unacknowledged client cleanup respectively: second, third and first columns. The
first draft build rejected an undeclared owner emphasis; the corrected briefs
give each accent its actual responsibility. No shared renderer changes.

Twelve pinned-source VM characterization tests passed for HTTP replay, cache-age
edge cases, scoped clearing, library retry/order/replay and storage failure.
These do not run browser IndexedDB or a deployed adapter. The
[delivery receipt](../verification/v1-offline-2026-09-08/README.md) records
final rendered capture, PDF, narration and publication inputs.

The first final guide review caught recovery-state names that were present only
in metadata, while the visible headings used different phrases. The document
fallback headings now explicitly name network failure, exact cache miss and home
cache miss. The complete final matrix is rerun against those corrected labels;
narration text is unchanged. The failed report remains in the receipt.

## Delivery checkpoint before integration

The final candidate has 146 passing visual/accessibility checks and all three
lost-content controls rejected. The center completed all fourteen focused
navigation/playback/PDF checks, preserving 61 prior guide HTML files and four
shared UI files byte for byte. All 597 prior recordings are unchanged; the
twenty new tracks add 1,819.373 seconds. Model optical review inspected every
final desktop slide, the full PDF sequence, native PDF pages 6, 7, 11 and 14,
and mobile freshness, queue handoff and record specimens. Normal-speed human
listening and willing-reader acceptance remain unperformed.

The final browser run binds formatted canonical JSON. Its rendered HTML and all
80 capture hashes match the visually inspected prior passing run exactly; the
prior report remains retained with that boundary. The original coverage attempt
was correctly rejected where supporting slides cited code but did not cite the
assigned documentation unit. The final citations and assembled fingerprints now
name the governing unit explicitly, and all four assignments pass.

Parallel Eve computer-use provider work at `a0ab808324` changes no presentation
files or cited offline implementations. Integrate it when publishing this
package. The broader V1 review continues with Tara practice/session and
workbench depth.

## Integration after the remote presentation update

The first push was rejected because origin/main advanced to
`949d68e86ef6477118d446e7f5c9d2e0b09aa4dc`, including three Eve operating-state
slides and their reviewed assignment. Preserve that published guide, audio and
PDF. The canonical catalog merges both additions; generated catalog, coverage,
duration and narration metadata are rebuilt from those combined inputs. The
offline guide HTML and twenty recordings remain unchanged. The final center
baseline is this integration commit, containing 600 prior recordings and 61
prior guide HTML files; the original authoring baseline remains 85f3a7.
Integration-specific center, browser-input, assignment and changed-contract
checks bind the combined 620-slide library. Pre-integration reports retain their
original scope and hashes in the receipt’s attempts directory.

The integration check preserves sixty browser captures and every PDF page render
exactly. Only the twenty expanded sidebar captures change their Eve/library
duration text; the offline slide region remains identical. The updated sidebar
was inspected. The exact approved-addition set in the V2 rollback preservation
test now includes the twenty reviewed offline slide IDs while retaining equality
checks for all 597 baseline slide and narration records. All 31 affected
integration tests pass after that explicit catalog update.
