feat(context): doc catalog — frontmatter/H1 descriptor as read-on-demand index

Docs were indexed as up to 40 h1-h3 headings per .md (a TOC flood) and
force-fed or excluded wholesale by a binary docs gate. Replace with a
compact, always-on catalog the agent reads on demand.

- RepoMapIndexer: .md now yields ONE descriptor (frontmatter
  description/summary/title -> first # H1 -> first prose line), not a
  heading dump. Replay-safe (still List<String> symbols).
- SessionOrchestrator: always-on '## Docs available (file_read to open)'
  catalog — top-N docs by recency as 'path — descriptor', hard-capped at
  DOCS_CATALOG_MAX. One line each, so it stays present without the
  poisoning the docs gate suppressed. Source layout still excludes doc
  content for non-doc prompts.

apps/server/src/main/kotlin/com/correx/apps/server/memory/RepoMapIndexer.kt

@@ -59,20 +59,58 @@ class RepoMapIndexer(
     }
 
     private fun extractSymbols(path: Path): List<String> {
-        val regex = SYMBOL_PATTERNS[path.extension] ?: return emptyList()
+        // Docs are indexed as a single "what is this" descriptor, not a table-of-contents dump: a
-        val text = runCatching { path.readText() }.getOrElse { return emptyList() }
+        // heading flood (40 h1–h3 per doc across a large docs/ tree) buries real source and force-feeds
-        return regex.findAll(text)
+        // content the stage may never need. One descriptor lets the agent decide whether to file_read
-            .mapNotNull { it.groups["name"]?.value }
+        // the doc (2026-07-07 doc-injection rework). Non-doc source keeps top-level symbol extraction.
+        if (path.extension.equals("md", ignoreCase = true)) return listOfNotNull(docDescriptor(path))
+        val regex = SYMBOL_PATTERNS[path.extension]
+        val text = regex?.let { runCatching { path.readText() }.getOrNull() }
+        return text?.let {
+            regex.findAll(it)
+                .mapNotNull { m -> m.groups["name"]?.value }
                 .distinct()
                 .take(MAX_SYMBOLS_PER_FILE)
                 .toList()
+        } ?: emptyList()
     }
 
+    /**
+     * One-line "what is this doc about" descriptor, preferred in order: YAML frontmatter
+     * `description`/`summary`/`title`, then the first `# H1` heading, then the first prose line.
+     * This is the read-on-demand index the doc catalog surfaces — never the doc body.
+     */
+    private fun docDescriptor(path: Path): String? {
+        val lines = runCatching { path.readText() }.getOrNull()?.lines() ?: return null
+        val raw = frontmatterDescriptor(lines)
+            ?: lines.firstOrNull { it.trimStart().startsWith("#") }
+                ?.trimStart()?.trimStart('#')?.trim()?.takeIf { it.isNotBlank() }
+            ?: lines.firstOrNull { it.isNotBlank() && !it.startsWith("---") }?.trim()
+        return raw?.truncateDescriptor()
+    }
+
+    private fun frontmatterDescriptor(lines: List<String>): String? {
+        val hasOpen = lines.firstOrNull()?.trim() == "---"
+        val end = if (hasOpen) lines.drop(1).indexOfFirst { it.trim() == "---" } else -1
+        if (!hasOpen || end < 0) return null
+        return FRONTMATTER_KEYS.firstNotNullOfOrNull { key ->
+            lines.subList(1, end + 1)
+                .firstOrNull { it.trimStart().startsWith("$key:", ignoreCase = true) }
+                ?.substringAfter(':')?.trim()?.trim('"', '\'')
+                ?.takeIf { it.isNotBlank() }
+        }
+    }
+
+    private fun String.truncateDescriptor(): String =
+        if (length <= MAX_DESCRIPTOR_CHARS) this else take(MAX_DESCRIPTOR_CHARS).trimEnd() + "…"
+
     companion object {
         /** File extensions considered for indexing — reused by [WorkspaceStateProbe] for fingerprinting. */
         val INDEXED_EXTENSIONS: Set<String> get() = SYMBOL_PATTERNS.keys
 
         private const val MAX_SYMBOLS_PER_FILE = 40
+        private const val MAX_DESCRIPTOR_CHARS = 120
+        private val FRONTMATTER_KEYS = listOf("description", "summary", "title")
 
         // Top-level declarations only — best-effort per language. Bodies/locals are never matched.
         val SYMBOL_PATTERNS: Map<String, Regex> = mapOf(

apps/server/src/test/kotlin/com/correx/apps/server/memory/RepoMapIndexerTest.kt

@@ -112,27 +112,45 @@ class RepoMapIndexerTest {
     }
 
     @Test
-    fun `extracts markdown headings as symbols`(@TempDir root: Path) {
+    fun `markdown is indexed as a single H1 descriptor, not a heading dump`(@TempDir root: Path) {
         root.resolve("design.md").writeText(
             """
             # Game Design Document
             intro text with an inline # hash
 
             ## Combat System
-            body
-
             ### Damage Formula
-            #### Too Deep To Matter
             """.trimIndent(),
         )
 
         val entry = RepoMapIndexer().index(root).single()
         assertEquals("design.md", entry.path)
-        assertTrue(
+        assertEquals(listOf("Game Design Document"), entry.symbols)
-            entry.symbols.containsAll(listOf("Game Design Document", "Combat System", "Damage Formula")),
+    }
-            "missing headings in: ${entry.symbols}",
+
+    @Test
+    fun `markdown descriptor prefers frontmatter description over the H1`(@TempDir root: Path) {
+        root.resolve("spec.md").writeText(
+            """
+            ---
+            title: Ignored Title
+            description: The authoritative sessions module spec
+            ---
+            # Sessions Module
+            body
+            """.trimIndent(),
         )
-        assertFalse(entry.symbols.any { it.contains("Too Deep") }, "h4+ headings must not be matched")
+
+        val entry = RepoMapIndexer().index(root).single()
+        assertEquals(listOf("The authoritative sessions module spec"), entry.symbols)
+    }
+
+    @Test
+    fun `markdown descriptor falls back to first prose line when there is no heading`(@TempDir root: Path) {
+        root.resolve("notes.md").writeText("just a plain note with no heading at all")
+
+        val entry = RepoMapIndexer().index(root).single()
+        assertEquals(listOf("just a plain note with no heading at all"), entry.symbols)
     }
 
     @Test

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt

@@ -197,6 +197,11 @@ private const val REFERENCE_EXISTS_CODE = "REFERENCE_EXISTS"
 private const val OUTPUT_SUMMARY_LIMIT = 500
 private const val REPO_MAP_INJECT_TOP_K = 30
 private const val REPO_MAP_FILES_PER_DIR = 8
+// Read-on-demand doc catalog: the top-N docs (by repo-map recency score) surfaced as
+// `path — descriptor` so a stage learns which docs exist and can file_read one when relevant,
+// instead of docs being force-fed (or excluded outright). One line each, hard-capped, so it can
+// stay always-on without re-poisoning context (2026-07-07 doc-injection rework).
+private const val DOCS_CATALOG_MAX = 20
 
 // A stage prompt must explicitly ask about documentation for .md/docs paths to enter the repo
 // layout listing; otherwise docs only reach context through a real retrieval hit.
@@ -1698,9 +1703,12 @@ abstract class SessionOrchestrator(
         val recorded = eventStore.read(sessionId)
             .mapNotNull { it.payload as? RepoKnowledgeRetrievedEvent }
             .lastOrNull { it.stageId == stageId }
-        if (recorded != null) return repoEntriesOrMapFloor(sessionId, recorded.hits, stagePrompt)
+        if (recorded != null) {
+            return repoEntriesOrMapFloor(sessionId, recorded.hits, stagePrompt) + buildDocsCatalogEntry(sessionId)
+        }
 
-        val retriever = repoKnowledgeRetriever ?: return buildRepoMapEntries(sessionId, stagePrompt)
+        val retriever = repoKnowledgeRetriever
+            ?: return buildRepoMapEntries(sessionId, stagePrompt) + buildDocsCatalogEntry(sessionId)
         val hits = runCatching { retriever.retrieve(sessionId, stagePrompt, REPO_MAP_INJECT_TOP_K) }
             .getOrElse { e ->
                 if (e is CancellationException) throw e
@@ -1710,7 +1718,44 @@ abstract class SessionOrchestrator(
         // Record the raw retrieval for audit (a degenerate all-zero result is itself the signal that
         // the embedder is noop / the index is empty), but only useful hits ever reach the agent.
         if (hits.isNotEmpty()) emit(sessionId, RepoKnowledgeRetrievedEvent(sessionId, stageId, stagePrompt, hits))
-        return repoEntriesOrMapFloor(sessionId, hits, stagePrompt)
+        return repoEntriesOrMapFloor(sessionId, hits, stagePrompt) + buildDocsCatalogEntry(sessionId)
+    }
+
+    /**
+     * Always-on, capped catalog of the docs that exist, as `path — descriptor` lines (the descriptor
+     * is the doc's frontmatter/H1/first-line summary recorded on [RepoMapEntry.symbols]). This is the
+     * read-on-demand doc index: one line per doc lets a stage decide whether to file_read it, so docs
+     * stay discoverable without the heading-flood the docs gate was added to suppress. Distinct from
+     * [buildRepoMapEntries]/[repoEntriesOrMapFloor], which still exclude doc *content* from the source
+     * layout for non-doc prompts. Reads the recorded map (replay-safe — never re-scans the FS).
+     */
+    private suspend fun buildDocsCatalogEntry(sessionId: SessionId): List<ContextEntry> {
+        val map = eventStore.read(sessionId)
+            .mapNotNull { it.payload as? RepoMapComputedEvent }
+            .lastOrNull() ?: return emptyList()
+        val docs = map.entries
+            .filter { isDocPath(it.path) }
+            .sortedByDescending { it.score }
+            .take(DOCS_CATALOG_MAX)
+        if (docs.isEmpty()) return emptyList()
+        val content = buildString {
+            appendLine("## Docs available (file_read to open — do not assume contents)")
+            docs.forEach { entry ->
+                val descriptor = entry.symbols.firstOrNull()?.takeIf { it.isNotBlank() }
+                appendLine(if (descriptor != null) "- ${entry.path} — $descriptor" else "- ${entry.path}")
+            }
+        }.trimEnd()
+        return listOf(
+            ContextEntry(
+                id = ContextEntryId(UUID.randomUUID().toString()),
+                layer = ContextLayer.L3,
+                content = content,
+                sourceType = "docsCatalog",
+                sourceId = "docs-catalog",
+                tokenEstimate = estimateTokens(content),
+                role = EntryRole.SYSTEM,
+            ),
+        )
     }
 
     /**