SPB Git

Repository Public License

testing.md 31.3 KB

# Testing

Two layers:

  1. Engine tests — Swift Testing suites inside each package, run on the host Mac with swift test. They cover the logic that produces every finding, with no simulator involved.
  2. UI tests — 23 XCUITests that drive the real app in the Simulator with an isolated in-memory store.

There is a third layer nobody has run yet: a person with a physical device. Read What the tests do not prove before quoting any result.

# Running

bash
Scripts/build.sh kit                               # every package, debug, host Mac
cd Packages/ORIXNetworkLab && swift test           # one package
Scripts/build.sh uitest                            # UI tests on "ORIXSecure iPhone"
ORIX_SIM="ORIXSecure iPad" Scripts/build.sh uitest # UI tests on "ORIXSecure iPad"

# Engine tests: counts observed on 2026-11-21

swift test (debug) in each package, host Mac (macOS 27.0, Xcode 27.2, Swift 6.4.2), online. Counts are the "Test run with N tests" line of each run. Parameterized tests count once, whatever the number of arguments.

Package (test target) Tests Suites Result
ORIXSecureModel 5 1 passed
ORIXSecureEngines 33 6 passed
ORIXURLRisk 43 11 passed
ORIXFileInspection 96 9 passed
ORIXNetworkLab 89 14 passed
ORIXMediaCleanup 52 5 passed
ORIXReporting (ORIXReportingTests) 34 6 passed
ORIXReporting (ORIXWorkflowTests) 27 3 passed
Total 379 55

The same 379 tests also passed in a release build during the Scripts/benchmark.sh run of the same day. ORIXDesign has no test target.

# What each suite proves

# ORIXSecureModel (5)

Reading keeps "unavailable" explicit with its reason and survives a JSON round trip; findings sort with attention first, then by severity; ReportDocument round-trips; without a catalog, L() returns the English source text.

# ORIXSecureEngines (33)

  • SecurityAssessmentTests — a healthy device produces no attention finding and no claim beyond what was read; a missing passcode is high severity with evidence; unreadable signals become "not available", never a default; integrity indicators are labelled heuristic and listed; an outdated system is reported only against a dated reference; the summary counts states and its coverage figure is not a security level; JSON round trip.
  • ChecklistTests — the catalog is well formed; every answer maps to the right state (a "yes" is confirmed by the user, never verified); answers expire; optional items never become problems; the latest answer wins; every catalog item yields honest findings for every possible answer.
  • DeviceDiagnosticsTests — unavailable readings stay explicit; thermal and core findings; advice is grounded in readings; the system probe returns real values on the Mac running the test.
  • StorageAnalysisTests — capacity thresholds; the app's own storage is measured and its cleanup is verified; missing paths and single files.
  • PrivacyAuditTests — permission findings are facts about this app only; every permission kind and status is explained.
  • InvestigationTests — every symptom has a complete plan; with no evidence the "device compromise" hypothesis is always last and "less likely"; compromise needs a concrete sign; answers move hypotheses to "supported" with their basis; no hypothesis text claims detection.

# ORIXURLRisk (43)

  • Punycode (RFC 3492) — the section 7.1 vectors encode, decode and round-trip; malformed input is rejected; random strings round-trip and random garbage never crashes.
  • Scripts, confusables, edit distance, Public suffix subset, IP literals — script detection per label, skeletons, registrable-domain splits, obfuscated IPv4 forms (decimal, octal, hex) and IPv6.
  • Decomposition — a URL is split into scheme, host, port, path, query; international hosts show both forms; a default port is removed and is not a signal.
  • Scoring — level thresholds; the score is the capped sum of the signal weights; keywords alone stay "low"; every analysis states its limits.
  • Brands — no brand signal on any embedded genuine domain or its sub-domains; brand names inside ordinary words do not fire; a redirect to the same site or brand is not a signal.
  • Corpora — see below.
  • Link extraction, QR payloads — links in free text; Wi-Fi, contact, message, call and place payloads.
  • Robustness and timing — empty and odd inputs, random bytes; a 10 000-character input under 100 ms; mean time of typical links.

# The false-positive corpus

Tests/ORIXURLRiskTests/Corpora.swift holds 119 legitimate links — the last entries are known hard cases (real sites whose name contains a brand name) kept on purpose so the measured rate is honest — and 92 suspicious constructions (homographs and Punycode, brand look-alikes, IP hosts, userinfo/port/structure tricks, encodings, redirects and shorteners, downloads and TLDs, schemes and invalid input), each with the signals it must raise and a minimum level.

Printed by the run of 2026-11-21 (identical in debug and release):

text
BENIGN CORPUS: 119 links — no signals: 89, low: 28, elevated: 2, high: 0
BENIGN CORPUS: above low = 2 / 119 = 1.68 %
BENIGN CORPUS: above low: https://www.paypal-community.com/ [url.brand.embedded=35]
BENIGN CORPUS: above low: https://www.googleadservices.com/pagead/aclk?sa=L&ai=abc [url.brand.embedded=30]
SUSPICIOUS CORPUS: 92 links — no signals: 4, low: 31, elevated: 37, high: 20
  • Measured false-positive rate on this corpus: 1.68 % (2 of 119 benign links rated above "low"; none rated "high"). The test fails above 5 %, or if any benign link is rated "high".
  • The two misses are real: a brand name embedded in a domain the brand list does not know. They are kept visible rather than special-cased.
  • The suspicious test asserts that each link raises its expected signals and reaches its minimum level, and that every signal kind of the analyzer has at least one corpus entry. Four entries end at "no signals"; the test passes because their declared minimum level allows it.
  • This is a rate on a hand-built corpus of 119 links. It is not an estimate of the false-positive rate on the links people actually receive, and the corpus says nothing about detection of real phishing campaigns.

# ORIXFileInspection (96)

  • Magic-number detection — every detectable type has a minimal sample that is recognised, in memory and through a real file; ambiguous signatures are told apart (MZ without PE, CAFEBABE Mach-O vs Java class, UTF-16 BOM vs MP3 sync, ISO media brands, DMG trailer); display names avoid verdict words.
  • File names and extension mismatch — double extensions, right-to-left override, invisible characters, padding spaces, look-alike extension characters; a Windows program named .pdf is the high-severity case, a PNG named .jpg a low one; equivalent extensions are accepted.
  • Archives without extraction — ZIP described exactly from its directory; decompression bombs produce findings from declared sizes and nothing is inflated; overlapping entries, path traversal, encrypted flag, symbolic links, nested archives, ZIP64, truncated archives, lying end records; the walk stops at the entry limit (150 000 entries cut at 100 000); Office documents (macros, embedded and linked parts); TAR and GZIP headers.
  • PDF inspection — name tokens counted exactly, #xx-escaped names decoded, tokens across chunk boundaries counted once, large files scanned at both ends only and the report says so; PDFKit page count, links, locked PDFs.
  • Document and image structure — PNG/JPEG declared dimensions and trailing data, RTF embedded objects, OLE2 headers.
  • SHA-256 streaming — known vectors; a 1 MiB pattern matches the system shasum; chunk size does not change the result; cancellation stops the hash; a 64 MB file reaches the hasher in pieces of at most one chunk.
  • Duplicates, large files, removal — duplicates across nested folders with look-alikes that are not duplicates; hard links; only the listed files are removed and freed bytes are verified; a file in a read-only folder is reported as failed and stays in place; links are removed, never their target; a cancelled removal touches nothing further.
  • File inspector end to end — every sample through a real file, with hash, limits and honest wording; standing limits say what the app is not; a file larger than every scan window is read in bounded pieces.
  • Fuzzing — see below.

# ORIXNetworkLab (89)

  • X.509 parser — static fixtures; certificates generated with openssl at test time (RSA leaf, EC root, RSA-1024/SHA-1 self-signed, expired / not yet valid / expiring); a tampered self-signed certificate fails its signature check; malformed input (every truncation, wrong lengths, indefinite length, non-minimal length, nesting depth bomb) throws a typed error.
  • Host name matching (RFC 6125), Address classification, Latency statistics, DNS error mapping, Connectivity snapshots and change log, Proxy, VPN indicator and parsing helpers — pure logic tables.
  • Network / TLS / HTTPS findings — fixtures for ordinary home Wi-Fi, proxy, VPN heuristic, cellular Low Data Mode, offline, open Wi-Fi and WEP, unavailable Wi-Fi; healthy chain, expired, weak self-signed, untrusted issuer; HSTS and upgrade; every finding states what an app cannot see.
  • Local TLS server (loopback) and Live network — see Live network tests.

# ORIXMediaCleanup (52)

  • Perceptual fingerprints — deterministic and well formed; measured distances per transformation; noise images are unrelated; degenerate images do not crash; grouping precision and recall on 360 generated images (figures in performance.md).
  • Cleanup analyzer — a slowly drifting sequence is not chained into one group; capture sessions; lookback; the global pass is off by default; colour and luminance reject same-structure photos; keeper rules; exact duplicates by content hash are measured and pre-selected, probable duplicates are heuristic and never pre-selected; recoverable bytes are "unavailable" when sizes are unknown; deterministic under shuffling; cancellation.
  • Fingerprint cache — miss/store/hit across loads; a changed modification date invalidates; corrupt files (garbage, truncated, bit flip, empty, bad magic) are discarded, never trusted; version bump.
  • Selection, plan, deletion and report — only extra copies of exact duplicates are proposed by default; deletion removes exactly the selected identifiers and is verified; a declined system confirmation, a library error and a failed verification never count items as deleted; limited access is reported; the report states "Recently Deleted" and claims no freed-space amount. These run against a fake library that implements the same protocol as the Photos-backed one.
  • Performance — throughput and scaling assertions.

# ORIXReporting (34) and ORIXWorkflow (27)

  • Markdown, Plain text, JSON — every field of every finding appears; hostile user data is escaped; tables stay well formed; 80-column wrapping; JSON round-trips to an equal document and rejects foreign or newer files.
  • PDF — text is selectable and complete (read back with PDFKit); metadata; paper sizes and margins; a 300-finding document paginates with correct totals; a finding title stays with its state line; concurrent rendering.
  • Export, Redaction — safe file names, no overwriting; device name, IP addresses, network identifiers, user names in paths and URL queries can be redacted; no option changes nothing.
  • Workflow validation — a destructive step must ask for confirmation and must be preceded by its review step; decoding an invalid definition fails.
  • Workflow runner — event order; a declined confirmation never runs the destructive step; a destructive step is skipped when its review did not succeed; cancellation during a step, between steps and while waiting for confirmation never deletes; step time limit; parallel runs do not share context.
  • Workflow runs and library — every built-in workflow is valid and runs to the end.

# Cross-cutting assertions

# Honesty: forbidden words and required evidence

Each engine's tests pass every finding they produce through a shared check. A finding fails the build if:

Rule Where
it has no evidence (label and value) all engines (in ORIXSecureEngines: for verified, attention and user-confirmed states)
its explanation is missing or too short all engines
it is an attention finding without a severity, or a non-attention finding with one ORIXSecureEngines, ORIXURLRisk
a network finding does not state what was not inspected ORIXNetworkLab
heuristic confidence and heuristic verification disagree ORIXFileInspection
its text contains a forbidden word all engines, lists below

Forbidden words, as written in the tests:

Test target List
ORIXSecureEngines (Fixtures.swift) "100% secure", "malware-free", "no malware detected", "virus", "infected", "hacked", "your device is safe", "is safe to"
ORIXURLRisk (AnalyzerTests.swift) "safe", "unsafe", "safely", "malicious", "clean", "infected", "virus", "viruses"
ORIXFileInspection (TestSupport.swift) "safe", "malicious", "infected", "virus", "malware detected", "clean"
ORIXNetworkLab (TestSupport.swift) "hacked", "safe", "secure network"

The check covers titles, explanations, recommendations, "not inspected" statements and evidence labels (and, for files, evidence values), plus summaries, methodology texts and type display names where they exist. It runs over the whole benign and suspicious corpora, every checklist item with every answer, and every file sample.

The UI tests repeat the principle on screen: no "100%", no "No malware", no "is safe", no "°C" invented for a processor temperature, no button containing "boost".

# Fuzzing

All fuzzers use a seeded generator, so a failure is reproducible.

Target Input Assertion
Every file parser 2 000 random buffers no crash, no hang
Every file parser bit-flipped and truncated valid samples no crash; findings or typed errors
ZIP directory every header field randomised no crash; bounded walk
File-name analyzer random Unicode no crash
Public file inspector mutated samples written to real files no crash
X.509 / DER 2 000 random buffers no crash
X.509 / DER 4 000 mutations of a valid certificate 1 896 rejected with a typed error, 2 104 still structurally valid, 0 crashes (run of 2026-11-21)
X.509 / DER every truncation of a valid certificate each throws
Punycode random strings and garbage round trip or rejection, no crash
URL analyzer random bytes, twelve 10 000-character shapes, a 300 000-character input no crash; under 100 ms

This is mutation and random fuzzing inside unit tests. No coverage-guided fuzzer (libFuzzer) and no sanitizer run (ASan, TSan, UBSan) has been used.

# Live network tests

Packages/ORIXNetworkLab/Tests/ORIXNetworkLabTests/LiveTests.swift has two serialized suites.

  • Local TLS server (loopback) starts openssl s_server on a loopback port with a certificate generated for the test. It proves that report-only mode completes the handshake and reports the untrusted certificate, that enforcing mode aborts but still reports the certificate, that a name the certificate does not cover is reported, that a closed port fails fast with a readable reason, and it measures loopback latency. It needs no Internet. Without openssl each test prints LIVE skipped: … and returns.
  • Live network (skipped when offline) first asks Live.isOnline() — a satisfied network path and a successful resolution of apple.com within 4 s. When that is false, the tests that need the Internet print LIVE skipped: offline — … and return without asserting. When true, they inspect apple.com:443, connect to its address without SNI to obtain a real name mismatch, resolve an existing and a non-existing name, measure latency, follow the HTTP → HTTPS upgrade and browse Bonjour for 2 s.
  • Tests that do not depend on reachability always run: DNS timeout and cancellation, a blackhole address that must time out within its budget, an unresolvable host.

Every live test prints LIVE ran: … or LIVE skipped: …, so the log shows what really executed. A skipped live test is reported by Swift Testing as passed: on an offline machine the count stays 89, and only the LIVE lines tell the difference. On 2026-11-21 the machine was online and every live test printed LIVE ran.

# UI tests

UITests/ORIXSecureUITests.swift, class ORIXSecureUITests, 23 tests. Each launches the app with -ORIXUITest YES -ORIXReset YES (in-memory store, separate defaults, animations off), in English (-AppleLanguages (en)), light appearance unless stated; on iPad the tests run in landscape. Elements are found by accessibility identifier. Screenshots are attached to the result bundle.

Status: the result of the last run on the "ORIXSecure iPhone" and "ORIXSecure iPad" simulators (iOS 26.5) is recorded in RELEASE-REPORT.md.

# Test What it verifies
1 testFirstLaunchExplainsLimitsThenShowsTheOverview The welcome screen states "No app can scan an iPhone for malware"; after Start the overview appears with no alert and no system permission prompt at launch.
2 testOverviewShowsCountsNotAScore The overview says it shows what was checked, not how secure the device is; no "100%", no "No malware"; the methodology sheet says no security score is computed; the limits card lists "System-wide malware scan".
3 testAuditFindingShowsEvidenceMethodAndLimits A finding of the security audit opens a detail sheet with "What it means", "How it was established" and "Evidence"; the audit can be saved as a report.
4 testChecklistAnswerIsRecordedAsReportedByTheUser "0 of 16 checked" → answering Yes gives "1 of 16 checked" and a row labelled "Confirmed by you", never "Verified"; answering No turns it into "Attention recommended" with a recommended action.
5 testPrivacyAuditReadsOnlyItsOwnPermissions The privacy audit states that an app can only read its own permission states; the photo-permission finding says other apps' permissions cannot be read.
6 testInvestigationSeparatesObservationsReportsAndHypotheses "Suspicious link" investigation: answering that a password was entered makes the phishing hypothesis "Supported" while device compromise stays "Less likely"; "None is a confirmed diagnosis"; advice to change the password; the report is saved.
7 testStorageShowsMeasuredCapacityAndVerifiedOwnCleanup Volume capacity is shown; the limits card says another app's cache cannot be deleted.
8 testFileCleanupDeletesOnlySelectedDuplicatesAfterConfirmation On a real sample folder: "2 extra copies in 1 group", 5 files ranked, nothing selected by default (no Delete button); after "Select Extra Copies" and the confirmation "Delete 2 File(s)", the result says 2 files deleted.
9 testPhotoCleanupFindsDuplicatesAndDeletesOnlyAfterTwoConfirmations Goes through the real system photo-permission prompt (Allow Full Access), scans the seeded library, opens the exact-duplicates category (a "Keep" marker is shown), opens the review ("Recently Deleted" is stated), taps Delete, then declines the real iOS delete confirmation: the result says nothing was deleted.
10 testPhotoDeletionIsVerifiedAndReported Same path, but accepts the real iOS delete confirmation: generated photos are really deleted and the result says they were moved to Recently Deleted.
11 testLinkAnalyserExplainsALookalikeWithoutOpeningIt https://apple.com.account-verify.example.net/… is not rated "No signals" and its anatomy shows the registrable domain example.net; then https://www.apple.com/iphone/ typed by hand shows "not proof that a link is harmless" and never "is safe".
12 testPunycodeLookalikeShowsBothForms https://xn--pple-43d.com/id shows the Punycode form and names it.
13 testFileScannerInspectsARealArchive A real ZIP named invoice.pdf.zip is identified as ZIP, hashed, summarised with findings to review; limits card; report saved.
14 testNetworkLabRunsLiveProbes Runs the live probes from the Simulator: a TLS section appears, a certificate is read, "TLS 1." is shown, the certificate sheet shows a SHA-256 fingerprint. Needs the Mac to be online.
15 testDeviceListsUnavailableMeasurementsInsteadOfInventingThem In the Simulator the battery reading is "Not available"; the unavailable list names battery cycle count and processor temperature; no "°C" anywhere.
16 testBatteryAdvisorOffersNoBoosters The "not offered" section says there is no "RAM boost"; no button contains "boost".
17 testReportsCanBeReadAndExportedInFourFormats Three demo reports are listed; a report states "It is not a malware scan"; PDF, Markdown, JSON and plain text export options exist; exporting PDF opens the system share sheet.
18 testWorkflowRunsRealStepsAndSavesItsReport The built-in "Security check-up" workflow runs to the end; its first step is "Done".
19 testWorkflowEditorRefusesDeletionWithoutReview A new workflow containing only a photo-deletion step cannot be saved: an error is shown.
20 testSettingsAndCapabilityStatement Settings open; the capability statement says "ORIX Secure is not an antivirus".
21 testDarkAppearance The overview renders with the device in dark appearance (screenshot kept).
22 testEveryTopLevelScreenOpens Each of the 16 destinations launches to a navigation bar with the app in the foreground (on iPad, one screenshot per destination).
23 testFrenchCanadianLocalization Launched with -AppleLanguages (fr-CA): an interface title and an engine state title are in French, and an English title is absent; screenshots of the overview and the audit in French.

Notes on the two photo tests (9 and 10):

  • Scripts/seed-photos.sh runs before every build.sh uitest. It adds generated images (byte-identical copies, resized and brightened variants, unrelated pictures) and resets the app's photo permission, because simctl privacy grant photos has no effect on this runtime — the permission is granted by tapping the system alert, found through the SpringBoard process with English and French button labels.
  • Deletion uses the real PHPhotoLibrary change request, so the confirmation is the system's own dialog. The app cannot bypass it and neither can the test.
  • Both tests call XCTSkipUnless when the library has no exact duplicates (for instance when test 10 runs after a previous deletion without re-seeding). A skip is not a pass; check the log for skipped tests.
  • Only generated images are ever in that library.

Reading a run:

bash
python3 Scripts/shots.py ORIXSecure_iPhone                       # named screenshots
python3 Scripts/shots.py ORIXSecure_iPhone --hierarchy testName  # accessibility hierarchy at a failure
open build/results/uitest-ORIXSecure_iPhone.xcresult

# What the tests do not prove

  • No physical device was used. Everything ran on a Mac (engines) or in the iOS Simulator (app). In particular, never exercised:
    • camera QR scanning (the Simulator has no camera; only QR payload parsing is tested);
    • Face ID / Touch ID App Lock;
    • battery level, state and Low Power Mode readings (the Simulator reports "Not available", which is what test 15 asserts);
    • Wi-Fi name and security type (needs a device, the Wi-Fi entitlement at run time and location permission);
    • the Local Network permission prompt and Bonjour browsing from a device;
    • real photo libraries: iCloud Photos with originals not on the device, Live Photos, bursts, HEIC/RAW, tens of thousands of assets, limited-access selection;
    • the share extension invoked from another app;
    • performance, memory, heat and battery use on a device (see performance.md);
    • integrity indicators on real hardware, and behaviour on the oldest supported iOS version (18.0).
  • The engines were tested as macOS builds. Platform branches behind #if canImport(UIKit) or canImport(Photos) are compiled for iOS by the app build but exercised only through the UI tests.
  • Language. The Simulator's system language was French while the app ran in English (-AppleLanguages (en)). System dialogs therefore appeared in French and the app in English. One UI test launches the app in French (Canada) and checks two strings on two screens; Scripts/localize.py checks that every translation exists and that format specifiers match. Neither proves that the French reads well or fits on every screen: no French speaker reviewed all 2 000 strings in context, and the Settings names follow Apple's terminology from memory.
  • Accessibility. VoiceOver was not exercised by a person. The tests rely on accessibility identifiers and labels, which shows that elements are exposed, not that the reading order or wording is good. Dynamic Type at the largest sizes, Reduce Motion, Increase Contrast, Bold Text and Switch Control were not checked visually.
  • Layouts. One iPhone simulator and one iPad simulator (landscape). No small-screen iPhone, no iPad Split View or Slide Over, no portrait iPad run.
  • Corpora are small and hand-built. 119 benign and 92 suspicious links; file samples are minimal synthetic files. The suites prove the rules do what they say, not how the analyzers behave on real-world mail, archives or documents.
  • Live network tests prove behaviour against one site on one network on one day.
  • No sanitizers, no coverage figure, no long-running soak test.

# Manual test checklist for a physical device

Run on an iPhone (and repeat the starred steps on an iPad) with a Release or TestFlight build. Note the device model, iOS version and build number with the results.

Install and first launch

  1. Install the build; confirm the Home Screen icon in default, dark and tinted appearances (long-press the Home Screen → Edit → Customize).
  2. ★ First launch: the welcome screen states the limits; tapping Start shows the overview. No permission prompt appears until a feature asks for it.
  3. ★ Switch the device between light and dark appearance, then set Dynamic Type to the largest accessibility size: the overview, a finding detail and the checklist stay readable, nothing is truncated without recourse.

Security and privacy

  1. Security Audit: the passcode finding matches the device (remove the passcode on a test device and confirm it becomes a high-severity finding, then restore it).
  2. Security Checklist: answer one item Yes → "Confirmed by you"; answer No → "Attention recommended". Kill and relaunch the app: the answers persist.
  3. Privacy Audit: each permission row matches Settings → ORIX Secure. Change one in Settings, return: the row updates.
  4. Settings → App Lock: turn it on, background the app, return: Face ID (or Touch ID) is requested. Fail the biometric check on purpose: the fallback works and the content stays hidden. Check the app switcher snapshot.

Link Analyser and QR

  1. Paste a known look-alike link and a plain well-known link: signals and anatomy are shown; nothing is opened; no wording says a link is "safe".
  2. Scan a QR code containing a URL with the camera: the camera prompt appears with its purpose text, the payload is analysed, the link is not opened. Scan a Wi-Fi QR code: the network is described, not joined.
  3. Deny the camera permission and retry: a clear explanation and a route to Settings, no crash.
  4. From Safari, share a page to "Analyse Link": the extension shows the analysis inside the share sheet.

Files

  1. Files → open a PDF, a ZIP and an Office document through the File Scanner: type, SHA-256 and findings appear; compare one SHA-256 with shasum -a 256 on a Mac.
  2. Scan a file larger than 1 GB: progress advances, Cancel stops it promptly, the app stays responsive.
  3. File Cleanup on a folder with known duplicates: only the selected copies are deleted, after confirmation; the originals remain in Files.

Photo Cleanup

  1. First use: the system photo prompt appears. Choose "Limit Access" with a few photos: the app states that access is limited and scans only those.
  2. Grant full access on a real library; note the library size and the scan duration shown. Scan again: the second scan is visibly faster (cache). Cancel a scan midway: it stops promptly.
  3. Open "exact duplicates": each group shows which photo is kept. Review, confirm in the app, then decline the iOS dialog: nothing is deleted.
  4. Repeat and accept the iOS dialog: the photos are in Photos → Recently Deleted; the result screen says so and claims no freed-space amount.
  5. With iCloud Photos and "Optimise iPhone Storage" on: the scan does not trigger downloads (watch cellular/Wi-Fi data counters); items without a local thumbnail are reported as skipped.

Network Lab

  1. On Wi-Fi: connection type, interfaces and the live probes (DNS, latency, TLS, HTTPS) return values. Open a certificate: subject, issuer, validity and SHA-256 fingerprint are shown.
  2. Wi-Fi name and security type: the location prompt explains why; after allowing, the name and security type match Settings → Wi-Fi. After denying, the row says why it is unavailable.
  3. Local network scan: the Local Network prompt appears on first use; services announced on the LAN are listed. Deny the permission: the scan reports that it is unavailable.
  4. Switch to cellular, then to Airplane Mode: the change log records each transition; probes fail with readable reasons, not spinners. Turn on a VPN: the VPN indicator is shown as a heuristic.

Device and battery

  1. Device: model, iOS version, storage and thermal state are plausible; anything unavailable says why.
  2. Battery: level and charging state match the status bar; plug and unplug the charger; toggle Low Power Mode. No cycle count or temperature is invented.

Reports and workflows

  1. Save a report from the audit; export it as PDF, Markdown, JSON and plain text; open each in Files. Turn on redaction and confirm the device name and addresses are gone from the export.
  2. Run the "Security check-up" workflow to the end; cancel another run midway. Try to save a workflow with a deletion step and no review step: it is refused.

System behaviour

  1. ★ VoiceOver: walk through the overview, one finding detail and the photo review. Every control is announced with its state in words (not only "button"); the summary ring is skipped and its counts are read from the text.
  2. ★ iPad: sidebar and detail column in both orientations, Split View at one-third width, Stage Manager resize, hardware keyboard navigation.
  3. Set the device language to Français (Canada): every screen and one exported report are in French, with no English left over and no truncated label. Then leave the app idle for an hour and check Settings → Battery: no background activity.

Record measured timings and memory from steps 13 and 16 in performance.md, with the device and library size.

Files main
Menu

Theme