# Browser Session and Terminal Selector Tests

Run from the repository root:

```sh
python3 tests/web_ui_session/run.py
```

Requires a host C compiler, Python 3, and Node with Fetch/Response/ReadableStream
support (Node 18+). All compiler outputs and rendered scripts are temporary; no
firmware build, generated assets, or device writes are performed.

The runner compiles production `src/web_ui.c` with HTTPD and vendored-asset data
doubles. It reuses the HTTPD stub text from `tests/web_login_ui/run.py`, without
importing/executing that runner. Node executes the actual C-rendered application
and inline asset-failure script, not a separately maintained implementation.

Coverage:

- Resource selection, NULL/invalid input, setter/send failure propagation,
  eight-header ceiling, no-store document/application, unchanged vendor caching,
  nosniff/no-referrer/frame denial, exact inline-loader CSP hash and login fallback.
- Session validation before initial/retried/restored connections; memory-only
  CSRF header and empty ticket/logout bodies; safe-text username/absolute expiry.
- 401 shutdown and navigation once; manual recovery on 403; bounded Retry-After
  display/backoff for capacity; network errors never assert successful logout.
- Confirmed 204 logout, lost response confirmed by session 401, uncertain logout,
  cancellation, explicit recovery, and stale session/ticket/status/logout/WS work.
- Pagehide/bfcache restoration, late response bodies, and superseded session checks.
- Existing writer controls, 1,024-byte binary input chunks, raw binary output,
  observer input gating, and explicit Disconnect pausing reconnect.
- Authentication/ticket response cap 512 bytes, existing status cap 3,072 bytes,
  15-second request deadline, single status request in flight, bounded retry delay,
  and unchanged 5,000-line terminal scrollback.
- Admin-only navigation, explicit protected-ticket open, 20 switches preserving
  serial socket/client/writer IDs, separate output and selected-only input.
- Independent 64 KiB pending output limits, visible dropped-byte counters and
  resumed hidden output draining; 512-byte admin input frames and overflow close.
- Admin close/reopen/remote exit isolation, 401/logout/expiry/pagehide cancellation,
  late tickets/callbacks, handshake timeout, selected resize and three listener
  cleanup/restore cycles.
- Session identity changes (username, role or session-stable CSRF) force a clean
  document before view adoption, close old admin and prevent replacement-session
  logout; same-session restore retains both scrollbacks behind validation gating.
- Undefined initial dimensions recover at unchanged bounds; failed fits never
  populate the cache, readiness retries stop at three, and teardown fences stale
  callbacks even after restore.
- Read-only admin Settings/Serial: 10 selection cycles preserve both sockets/IDs,
  drain hidden output and block terminal input; exact eight-field/256-byte schema,
  explicit refresh, error/timeout containment, late cancellation/restore/identity,
  and concurrent serial reconnect without superseding admission.
- 8D.9 typed Serial actions: draft validation, Reset-only confirmation, bounded
  JSON/CSRF, explicit working/persisted effects, automatic bounded result checks,
  failure/timeout containment, session fencing and preserved sockets/writer identity.
- Lost-acknowledgement and replaced-result uncertainty survives repeated result
  checks, failed reads, refresh and navigation; a newly acknowledged explicit
  submission starts a new result context.
- Immediate completion on the first GET, pending then completion, 10-attempt
  exhaustion and manual recovery, 15-second overall abort during fetch/body reads,
  delayed timers/replies, automatic read errors, visible stale snapshots during
  refresh and after refresh failure for every terminal outcome, late refresh
  cancellation, no routine confirmations and Reset cancellation. Navigation,
  pagehide/restore, logout, expiry and changed identity cancel checks without
  automatic resumption.
- Repeated current Settings selection is a no-op during submission, between and
  during result checks, and during completion refresh: requests, timers, visible
  values/control state, final outcome and socket/writer identity remain intact.
  **35 Node groups total.**

## Automatic result-check budget

After a valid POST acknowledgement, the UI waits **1,000 ms** before the first
result GET and between completed pending-result checks. It makes **at most 10
GET attempts** and uses an independent **15,000 ms overall deadline**, measured
with the monotonic browser clock from acknowledgement. Each attempt first
revalidates the session; that time is included in the deadline. There is only
one automatic check in flight. Delayed timer callbacks and replies also check
this deadline. Expiry actively aborts the in-flight request and releases the UI
for manual recovery; late completions cannot update the view.

The first limit reached stops automatic checking. A read error also stops it.
The budget does not cancel backend work and is not a server execution deadline.
**POST is never automatically retried.** Lost acknowledgement requires explicit
Check Result recovery; manual checks do not restart automatic polling. Exhausted
or cancelled polling never resumes on navigation or bfcache restoration.

Every known terminal result, including failure/cancellation, triggers one working
snapshot refresh while retaining the operation outcome and any uncertainty
warning. Snapshot refresh is outside the auto-check budget and retains the
existing 15-second per-request bound (session validation and snapshot GET are
separate requests). Settings remain visible but conflicting controls are disabled
during work; old snapshots are explicitly stale during pending/uncertain work or
a failed refresh. A successful refresh replaces the browser draft. Only Reset
asks for confirmation, specifically because it overwrites saved configuration.

Tests use a deterministic clock and individually fired timer callbacks, including
callbacks invoked after cancellation and fetch/body doubles that ignore abort.
These deliberately exercise fences beyond normal browser cancellation behavior.

## Integration and known gaps

This covers 8D.3 session behavior, the 8D.6 selector, 8D.8 Settings and the 8D.9
Serial UI. Operation responses are fetch doubles, not end-to-end execution of
`web_serial_settings.c`, dispatcher work, serial reconfiguration or NVS persistence. The renderer
still relies on its caller to authenticate resources; protected asset failures
must be 401, never a redirect to HTML served as JavaScript. No Basic fallback is
implemented here. Existing 8D.5 server authorization/protocols are unchanged.

These tests model DOM, timers, fetch cancellation and WebSocket events. They do
not prove real-browser CSP enforcement, script-loading errors, TLS/HTTPD behavior,
actual bfcache policy, cookie expiry, server revocation, or hardware serial byte
integrity, actual xterm escape parsing, hidden prompts, or desktop/mobile layout.
Prior 8D.6 signoff stands; the new Settings build and pending target checklist are in
`docs/phase8d8_implementation.md`. No target resource reserve is claimed. Browser secret
references are dropped and never persisted/logged, but JavaScript cannot securely
wipe engine-managed strings.