Stop reader mode passing a site's cookie notice off as the article

Reported live: reader mode on a New Republic feature showed nothing but
the cookie disclaimer, and had cached it.

Reproduced outside the app by running ReaderExtractor's own injected
script (Readability.js and all) against the real page. With the article
body present, extraction is perfect. With the body absent from the DOM at
the moment the script runs, Readability's first pass -- which discards
"unlikely candidates", and the Didomi banner's class is
didomi-notice-banner -- comes back under its own 500-character threshold,
so it retries with that stripping switched off. The banner's single
paragraph is 510 characters, i.e. just over the threshold, so it comes
back as the article with the right title attached.

Nothing on the Kotlin side looked at that. Any non-blank content counted
as ExtractionResult.Success, was written to articles.extractedContent,
and was never reconsidered -- so one bad load became a permanent cookie
disclaimer, which is the "extractedContent cached: true" in the report.

ConsentNotice recognises such a result: a hard plain-text length cap
first (nothing long is ever rejected, however much consent boilerplate it
quotes), then either a known CMP container id -- Readability strips
classes but keeps ids -- or two or more distinct consent phrases.

The accept/retry/give-up decision moves out of the WebView callback into
decideEvalOutcome so it can be tested without a device. A consent notice
is now "not ready yet", which routes it into the retry the extractor
already does for an empty result; if it survives all six attempts that is
what the page serves, and it fails as ExtractionFailure.ConsentInterstitial
rather than being accepted.

A fixed extractor cannot repair what is already on disk, so
loadCachedExtraction drops a cached notice on read and re-extracts, in
both ReaderScreen and LinkReaderScreen. No schema change: the repair
happens when the article is next opened. Reader mode is still never
turned on heuristically -- replacing a discarded notice does not flip it.

The bug report now also carries the last extraction error, without which
a report about bad reader content cannot say whether extraction failed.

Tests: 59 -> 78.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018AwnALZL6bD4VBtATDVv3B

3 files changed, 296 insertions, 0 deletions

diff --git a/app/src/test/java/org/terst/feedreader/data/CachedExtractionTest.kt b/app/src/test/java/org/terst/feedreader/data/CachedExtractionTest.kt
new file mode 100644
index 0000000..72dc73f
--- /dev/null
+++ b/app/src/test/java/org/terst/feedreader/data/CachedExtractionTest.kt
@@ -0,0 +1,67 @@
+package org.terst.feedreader.data
+
+import kotlinx.coroutines.test.runTest
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertFalse
+import org.junit.Assert.assertNull
+import org.junit.Assert.assertTrue
+import org.junit.Test
+
+/**
+ * The reported bug came with "extractedContent cached: true", which is the half
+ * that a fixed extractor cannot reach: the cookie disclaimer was already written
+ * to `articles.extractedContent`, and every open after that read it straight back
+ * out without ever looking at it again.
+ *
+ * So a poisoned cache has to be recognised and dropped on read, not just kept out
+ * on write.
+ */
+class CachedExtractionTest {
+
+    private val cookieNotice =
+        "<div id=\"readability-page-1\" class=\"page\"><div id=\"didomi-host\"><p>This website uses " +
+            "cookies and other tracking technologies (“Cookies”) to enhance user experience, to " +
+            "display targeted ads, and to analyze performance and user activity on our website. For " +
+            "more information about our use of Cookies and our privacy practices, see our Privacy " +
+            "Policy. By continuing to use this website, you agree to our Terms and Conditions.</p>" +
+            "</div></div>"
+
+    @Test
+    fun `real cached content is used as-is and the cache is left alone`() = runTest {
+        var cleared = false
+        val content = "<p>" + "The actual article body. ".repeat(60) + "</p>"
+        val result = loadCachedExtraction(content) { cleared = true }
+        assertEquals(content, result.content)
+        assertFalse(result.discardedConsentNotice)
+        assertFalse("a good cache must not be cleared", cleared)
+    }
+
+    @Test
+    fun `a cached consent notice is discarded and deleted`() = runTest {
+        var cleared = false
+        val result = loadCachedExtraction(cookieNotice) { cleared = true }
+        assertNull("the notice must not be shown as the article", result.content)
+        assertTrue(result.discardedConsentNotice)
+        assertTrue("the poisoned row must be cleared so a re-extract can replace it", cleared)
+    }
+
+    @Test
+    fun `nothing cached touches nothing`() = runTest {
+        var cleared = false
+        val result = loadCachedExtraction(null) { cleared = true }
+        assertNull(result.content)
+        assertFalse(result.discardedConsentNotice)
+        assertFalse(cleared)
+    }
+
+    // A blank value is as good as nothing cached -- it is not a consent notice, so
+    // reporting it as a discarded one would make the caller re-extract on a false
+    // premise.
+    @Test
+    fun `a blank cached value is not reported as a discarded notice`() = runTest {
+        var cleared = false
+        val result = loadCachedExtraction("   ") { cleared = true }
+        assertFalse(result.discardedConsentNotice)
+        assertFalse(cleared)
+    }
+}
diff --git a/app/src/test/java/org/terst/feedreader/readability/ConsentNoticeTest.kt b/app/src/test/java/org/terst/feedreader/readability/ConsentNoticeTest.kt
new file mode 100644
index 0000000..7f4e9cc
--- /dev/null
+++ b/app/src/test/java/org/terst/feedreader/readability/ConsentNoticeTest.kt
@@ -0,0 +1,140 @@
+package org.terst.feedreader.readability
+
+import org.junit.Assert.assertFalse
+import org.junit.Assert.assertTrue
+import org.junit.Test
+
+/**
+ * Reported live: reader mode on a New Republic feature article showed nothing but
+ * the site's cookie disclaimer, and cached it.
+ *
+ * Reproduced outside the app by running the extractor's own injected script
+ * (ReaderExtractor's `script`, Readability.js and all) against the real page: when
+ * the article body is present, extraction is perfect. When the body is NOT in the
+ * DOM at the moment Readability runs, Readability's own fallback pass kicks in --
+ * its first pass strips "unlikely candidates" (which the consent banner is, its
+ * class says `didomi-notice-banner`), comes back under its 500-character
+ * threshold, and so it retries with that stripping disabled. The Didomi notice's
+ * single paragraph is 510 characters, i.e. just over that threshold, so it is
+ * accepted as the article.
+ *
+ * [EXTRACTED_COOKIE_NOTICE] below is not a hand-written approximation of that --
+ * it is the verbatim `content` string that reproduction produced.
+ *
+ * Nothing on the Kotlin side looked at that: any non-blank `content` counted as a
+ * successful extraction, got written to `articles.extractedContent`, and was never
+ * re-checked. Hence a permanently cached cookie disclaimer.
+ */
+class ConsentNoticeTest {
+
+    @Test
+    fun `the reported New Republic extraction is recognised as a consent notice`() {
+        assertTrue(ConsentNotice.looksLikeConsentNotice(EXTRACTED_COOKIE_NOTICE))
+    }
+
+    // Readability strips class attributes but keeps ids, so the CMP container's id
+    // survives into the extracted HTML and is the most precise signal available.
+    @Test
+    fun `other consent-management platforms are recognised by their container id`() {
+        val cmpIds = listOf(
+            "didomi-host",
+            "onetrust-banner-sdk",
+            "onetrust-consent-sdk",
+            "CybotCookiebotDialog",
+            "qc-cmp2-container",
+            "sp_message_container_1234",
+            "truste-consent-track",
+            "usercentrics-root",
+            "osano-cm-dialog",
+            "iubenda-cs-banner",
+            "optanon-alert-box-wrapper",
+            "cookie-law-info-bar",
+        )
+        for (id in cmpIds) {
+            val html = """<div id="readability-page-1"><div id="$id"><p>Please review your choices.</p></div></div>"""
+            assertTrue("$id should be recognised as a consent container", ConsentNotice.looksLikeConsentNotice(html))
+        }
+    }
+
+    // Not every cookie banner is a known CMP. A short result whose text is
+    // unmistakably consent boilerplate counts too, with no id to go on.
+    @Test
+    fun `a short unbranded cookie banner is recognised from its wording`() {
+        val html = """
+            <div id="readability-page-1"><div><p>We use cookies and similar tracking technologies
+            to personalise content and ads. By continuing to use this site you consent to our use
+            of cookies. You can manage your preferences at any time.</p></div></div>
+        """.trimIndent()
+        assertTrue(ConsentNotice.looksLikeConsentNotice(html))
+    }
+
+    // One stray phrase is not enough -- plenty of ordinary writing mentions cookies.
+    @Test
+    fun `a single incidental phrase is not enough`() {
+        val html = """<div id="readability-page-1"><div><p>The recipe calls for cookies.
+            We use cookies in the frosting too.</p></div></div>"""
+        assertFalse(ConsentNotice.looksLikeConsentNotice(html))
+    }
+
+    // The whole point is to not throw away real articles. A genuine article ABOUT
+    // cookie-consent law would hit every phrase there is; length is what separates
+    // it from a banner. Nothing long is ever rejected.
+    @Test
+    fun `a long article about cookie consent is never rejected`() {
+        val paragraph = "<p>We use cookies and other tracking technologies, and by continuing " +
+            "to use this website you consent to our use of cookies -- that is the boilerplate " +
+            "the GDPR produced, and this article is about where it came from. You can manage " +
+            "your preferences, in theory.</p>"
+        val html = "<div id=\"readability-page-1\"><div>${paragraph.repeat(12)}</div></div>"
+        assertTrue("fixture must be past the length cap", html.length > 2_000)
+        assertFalse(ConsentNotice.looksLikeConsentNotice(html))
+    }
+
+    // Short posts are normal (kottke-style linkblogging is called out in
+    // ReaderExtractor's own comments) and must survive.
+    @Test
+    fun `an ordinary short post is not a consent notice`() {
+        val html = """
+            <div id="readability-page-1"><div><p>A lovely short note about a bridge in Lisbon,
+            with one link and nothing else. This is the entire post.</p></div></div>
+        """.trimIndent()
+        assertFalse(ConsentNotice.looksLikeConsentNotice(html))
+    }
+
+    @Test
+    fun `null and blank content are not consent notices`() {
+        assertFalse(ConsentNotice.looksLikeConsentNotice(null))
+        assertFalse(ConsentNotice.looksLikeConsentNotice(""))
+        assertFalse(ConsentNotice.looksLikeConsentNotice("   \n  "))
+    }
+
+    // Markup must not count towards the length budget: a banner wrapped in enough
+    // nested divs and inline styles to exceed the cap in raw characters is still a
+    // banner.
+    @Test
+    fun `length is measured on text not markup`() {
+        val padding = "<div style=\"${"border-color: rgb(0, 0, 0); ".repeat(80)}\">"
+        val html = padding + EXTRACTED_COOKIE_NOTICE + "</div>"
+        assertTrue("fixture must be past the length cap in raw characters", html.length > 2_000)
+        assertTrue(ConsentNotice.looksLikeConsentNotice(html))
+    }
+
+    companion object {
+        /**
+         * Verbatim output of running ReaderExtractor's injected script against the
+         * reported page in a state where the article body was absent. Do not tidy
+         * this up -- its value is that it is exactly what the app produced.
+         */
+        const val EXTRACTED_COOKIE_NOTICE: String =
+            "<div id=\"readability-page-1\" class=\"page\"><div id=\"didomi-host\" data-testid=\"notice\" " +
+                "data-nosnippet=\"true\" lang=\"en\" data-lang-dir=\"ltr\"><p>This website uses cookies and " +
+                "other tracking technologies (“Cookies”) to enhance user experience, to display " +
+                "targeted ads, and to analyze performance and user activity on our website. We also may " +
+                "share information about your activity on our website with our advertising and analytics " +
+                "providers. For more information about our use of Cookies and our privacy practices, see " +
+                "our <a href=\"https://newrepublic.com/pages/privacy\" target=\"_blank\">Privacy Policy</a>. " +
+                "By continuing to use this website, you agree to our " +
+                "<a href=\"https://newrepublic.com/pages/terms-and-conditions\" target=\"_blank\">Terms and " +
+                "Conditions</a>, including our dispute resolution provisions.</p></div></div>"
+    }
+}
diff --git a/app/src/test/java/org/terst/feedreader/readability/EvalOutcomeTest.kt b/app/src/test/java/org/terst/feedreader/readability/EvalOutcomeTest.kt
new file mode 100644
index 0000000..8995e22
--- /dev/null
+++ b/app/src/test/java/org/terst/feedreader/readability/EvalOutcomeTest.kt
@@ -0,0 +1,89 @@
+package org.terst.feedreader.readability
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+
+/**
+ * The decision ReaderExtractor makes about each `evaluateJavascript` result:
+ * accept it, inject again because the page probably was not ready, or give up
+ * with a reason. It used to live inline in a WebView callback, which meant the
+ * only way to exercise it was on a device against a live page.
+ *
+ * The case that mattered and was never covered: a result that is perfectly
+ * well-formed and non-blank, and is the page's cookie-consent banner rather than
+ * the article. That counted as a success, got cached, and stayed cached.
+ */
+class EvalOutcomeTest {
+
+    private val maxAttempts = 6
+
+    private fun decide(article: ExtractedArticle?, attempt: Int, raw: String = "\"{}\"") =
+        decideEvalOutcome(article, raw, attempt, maxAttempts)
+
+    private fun article(content: String?, error: String? = null) =
+        ExtractedArticle(title = "t", content = content, textContent = content, error = error)
+
+    @Test
+    fun `real content is accepted on the first attempt`() {
+        val a = article("<p>" + "The actual article body. ".repeat(60) + "</p>")
+        val decision = decide(a, attempt = 1)
+        assertEquals(EvalDecision.Done(ExtractionResult.Success(a)), decision)
+    }
+
+    // Retrying cannot change Readability's opinion of the markup, so a thrown
+    // error is final however early it arrives.
+    @Test
+    fun `a readability error is final immediately`() {
+        val decision = decide(article(null, error = "TypeError: x is null"), attempt = 1)
+        assertEquals(
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.ReadabilityError("TypeError: x is null"))),
+            decision
+        )
+    }
+
+    @Test
+    fun `an undecodable result retries and then reports a bad result`() {
+        assertEquals(EvalDecision.Retry, decide(null, attempt = 1, raw = "null"))
+        val last = decide(null, attempt = maxAttempts, raw = "null")
+        assertEquals(EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.BadResult("null"))), last)
+    }
+
+    @Test
+    fun `blank content retries and then reports no content`() {
+        assertEquals(EvalDecision.Retry, decide(article(""), attempt = 1))
+        assertEquals(
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.NoContent)),
+            decide(article(""), attempt = maxAttempts)
+        )
+    }
+
+    // A cookie banner is exactly what a half-loaded page yields -- the banner is
+    // injected early, the article body may not be in the DOM yet -- so the right
+    // first move is the retry the extractor already does for an empty result.
+    @Test
+    fun `a consent notice retries rather than being accepted`() {
+        val decision = decide(article(ConsentNoticeTest.EXTRACTED_COOKIE_NOTICE), attempt = 1)
+        assertEquals(EvalDecision.Retry, decision)
+    }
+
+    @Test
+    fun `a consent notice that survives every attempt is reported as a failure`() {
+        val decision = decide(article(ConsentNoticeTest.EXTRACTED_COOKIE_NOTICE), attempt = maxAttempts)
+        assertEquals(
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.ConsentInterstitial)),
+            decision
+        )
+    }
+
+    // The failure has to be distinguishable from "no content": one means the page
+    // has no article, the other means the page would not show it without consent.
+    @Test
+    fun `the consent failure says what happened`() {
+        val f: ExtractionFailure = ExtractionFailure.ConsentInterstitial
+        assertTrue(f.userMessage().isNotBlank())
+        assertTrue(f.diagnostic().isNotBlank())
+        assertTrue(f.userMessage().lowercase().contains("cookie") || f.userMessage().lowercase().contains("consent"))
+        assertTrue(f.userMessage() != ExtractionFailure.NoContent.userMessage())
+    }
+}