Fix empty section content + surface richer citations

internal/config/config_treewalk_test.go

@@ -18,13 +18,13 @@ func TestForwardTreeWalkMaxCitations(t *testing.T) {
 		os.Setenv("VLE_RETRIEVAL_TREEWALK_MAX_CITATIONS", prevVLE)
 	}()
 
-	// Engine default is 3 with no env set.
+	// Engine default is 6 with no env set.
 	os.Unsetenv("VLS_RETRIEVAL_TREEWALK_MAX_CITATIONS")
 	os.Unsetenv("VLE_RETRIEVAL_TREEWALK_MAX_CITATIONS")
 	cfg := Default()
 	applyEnvOverrides(&cfg)
-	if cfg.Engine.Retrieval.TreeWalk.MaxCitations != 3 {
-		t.Errorf("default max_citations = %d, want 3", cfg.Engine.Retrieval.TreeWalk.MaxCitations)
+	if cfg.Engine.Retrieval.TreeWalk.MaxCitations != 6 {
+		t.Errorf("default max_citations = %d, want 6", cfg.Engine.Retrieval.TreeWalk.MaxCitations)
 	}
 
 	// VLE_ alone forwards through.
@@ -43,12 +43,12 @@ func TestForwardTreeWalkMaxCitations(t *testing.T) {
 		t.Errorf("VLS_ should win: max_citations = %d, want 1", cfg3.Engine.Retrieval.TreeWalk.MaxCitations)
 	}
 
-	// Garbled value preserves the engine default (3), does not zero it.
+	// Garbled value preserves the engine default (6), does not zero it.
 	os.Unsetenv("VLS_RETRIEVAL_TREEWALK_MAX_CITATIONS")
 	os.Setenv("VLE_RETRIEVAL_TREEWALK_MAX_CITATIONS", "heaps")
 	cfg4 := Default()
 	applyEnvOverrides(&cfg4)
-	if cfg4.Engine.Retrieval.TreeWalk.MaxCitations != 3 {
-		t.Errorf("garbage value should preserve default 3, got %d", cfg4.Engine.Retrieval.TreeWalk.MaxCitations)
+	if cfg4.Engine.Retrieval.TreeWalk.MaxCitations != 6 {
+		t.Errorf("garbage value should preserve default 6, got %d", cfg4.Engine.Retrieval.TreeWalk.MaxCitations)
 	}
 }

pkg/config/config.go

@@ -760,7 +760,7 @@ func Default() Config {
 				Enabled:          true,
 				MaxHops:          8,
 				PageContentLimit: 16000,
-				MaxCitations:     3,
+				MaxCitations:     6,
 			},
 		},
 		Ingest: IngestConfig{

pkg/config/config_test.go

@@ -684,8 +684,8 @@ func TestTreeWalkDefaults(t *testing.T) {
 	if cfg.Retrieval.TreeWalk.PageContentLimit != 16000 {
 		t.Errorf("page_content_limit = %d, want 16000", cfg.Retrieval.TreeWalk.PageContentLimit)
 	}
-	if cfg.Retrieval.TreeWalk.MaxCitations != 3 {
-		t.Errorf("max_citations = %d, want 3", cfg.Retrieval.TreeWalk.MaxCitations)
+	if cfg.Retrieval.TreeWalk.MaxCitations != 6 {
+		t.Errorf("max_citations = %d, want 6", cfg.Retrieval.TreeWalk.MaxCitations)
 	}
 	if cfg.Retrieval.TreeWalk.Model != "" {
 		t.Errorf("model default should be empty (inherit), got %q", cfg.Retrieval.TreeWalk.Model)
@@ -802,8 +802,8 @@ func TestTreeWalkEnvOverrideRejectsBad(t *testing.T) {
 	if cfg.Retrieval.TreeWalk.PageContentLimit != 16000 {
 		t.Errorf("garbage page_content_limit env should preserve default, got %d", cfg.Retrieval.TreeWalk.PageContentLimit)
 	}
-	if cfg.Retrieval.TreeWalk.MaxCitations != 3 {
-		t.Errorf("garbage max_citations env should preserve default 3, got %d", cfg.Retrieval.TreeWalk.MaxCitations)
+	if cfg.Retrieval.TreeWalk.MaxCitations != 6 {
+		t.Errorf("garbage max_citations env should preserve default 6, got %d", cfg.Retrieval.TreeWalk.MaxCitations)
 	}
 }
 

pkg/parser/empty_sections_test.go

@@ -0,0 +1,88 @@
+package parser
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestFoldEmptyLeafSections proves the post-parse pass eliminates empty
+// "structural" leaf fragments — the multi-line display-title remnants and
+// running headers that otherwise ship as sections returning empty content
+// to the browser / API — while preserving their title text on the sibling
+// body they head, and never touching internal nodes (empty headings that
+// carry real sub-sections).
+func TestFoldEmptyLeafSections(t *testing.T) {
+	t.Parallel()
+
+	t.Run("empty leaf folds its title into the following sibling", func(t *testing.T) {
+		in := []Section{
+			{Level: 1, Title: "How Vectorless", Content: "", PageStart: 3, PageEnd: 3},
+			{Level: 1, Title: "Body", Content: "Vectorless walks the document tree.", PageStart: 3, PageEnd: 4},
+		}
+		out := foldEmptyLeafSections(in)
+		if len(out) != 1 {
+			t.Fatalf("want 1 section (empty leaf folded away), got %d: %+v", len(out), out)
+		}
+		if strings.Contains(out[0].Content, "walks the document tree") == false {
+			t.Errorf("body content lost: %q", out[0].Content)
+		}
+		if !strings.Contains(out[0].Content, "How Vectorless") {
+			t.Errorf("folded title not preserved in content: %q", out[0].Content)
+		}
+		if out[0].PageStart != 3 {
+			t.Errorf("PageStart should union to 3, got %d", out[0].PageStart)
+		}
+	})
+
+	t.Run("consecutive empty fragments accumulate onto one body", func(t *testing.T) {
+		secs := []Section{
+			{Level: 1, Title: "THE SYSTEM, END-TO-END", Content: "", PageStart: 2, PageEnd: 2},
+			{Level: 1, Title: "How Vectorless", Content: "", PageStart: 2, PageEnd: 2},
+			{Level: 1, Title: "actually works.", Content: "", PageStart: 2, PageEnd: 2},
+			{Level: 1, Title: "Intro", Content: "The system reads structure first.", PageStart: 2, PageEnd: 3},
+		}
+		out := foldEmptyLeafSections(secs)
+		if len(out) != 1 {
+			t.Fatalf("want 1 section, got %d: %+v", len(out), out)
+		}
+		for _, frag := range []string{"THE SYSTEM", "How Vectorless", "actually works", "reads structure first"} {
+			if !strings.Contains(out[0].Content, frag) {
+				t.Errorf("fragment %q missing from folded content: %q", frag, out[0].Content)
+			}
+		}
+	})
+
+	t.Run("empty internal node (heading with children) is preserved", func(t *testing.T) {
+		in := []Section{
+			{Level: 1, Title: "Chapter", Content: "", Children: []Section{
+				{Level: 2, Title: "Sub", Content: "real body", PageStart: 5, PageEnd: 5},
+			}},
+		}
+		out := foldEmptyLeafSections(in)
+		if len(out) != 1 || len(out[0].Children) != 1 {
+			t.Fatalf("internal node must survive with its child, got %+v", out)
+		}
+	})
+
+	t.Run("trailing empty fragment with no following body attaches to previous", func(t *testing.T) {
+		in := []Section{
+			{Level: 1, Title: "Body", Content: "content here", PageStart: 1, PageEnd: 1},
+			{Level: 1, Title: "Dangling Header", Content: "", PageStart: 2, PageEnd: 2},
+		}
+		out := foldEmptyLeafSections(in)
+		if len(out) != 1 {
+			t.Fatalf("want 1 section, got %d", len(out))
+		}
+		if !strings.Contains(out[0].Content, "Dangling Header") {
+			t.Errorf("trailing title lost: %q", out[0].Content)
+		}
+	})
+
+	t.Run("lone empty leaf with no siblings is kept (nothing to fold into)", func(t *testing.T) {
+		in := []Section{{Level: 1, Title: "Only", Content: ""}}
+		out := foldEmptyLeafSections(in)
+		if len(out) != 1 {
+			t.Fatalf("lone empty leaf should be kept, got %d", len(out))
+		}
+	})
+}

pkg/parser/pdf.go

@@ -424,10 +424,32 @@ func (p *PDF) parseDoc(_ context.Context, buf []byte) (*ParsedDoc, error) {
 			// depth relative to the font-derived level. We only ever DEEPEN
 			// (never change the base level), so top-level headings — numbered
 			// "3" or not ("Abstract") — stay siblings at the same font level.
+			newLevel := lvl
 			if nd, ok := numberedHeadingDepth(text); ok && nd > 1 {
-				lvl += nd - 1
+				newLevel += nd - 1
 			}
-			current = &flat{level: lvl, title: text}
+			// Multi-line display heading: a run of heading rows at the SAME
+			// level, on the SAME page, with NO body text accrued between them
+			// is one visual heading wrapped across lines — e.g. a hero title
+			// "THE SYSTEM, END-TO-END" / "How Vectorless" / "actually works."
+			// Left un-merged, each line becomes its own heading flat whose
+			// body is empty, shipping as a "purely structural" Section that
+			// returns empty content to the section browser / API. Fold the
+			// continuation line into the current heading instead so the body
+			// that follows attaches to the whole title and no empty fragment
+			// section is emitted. We only merge when the current flat is
+			// itself an as-yet-empty heading at the same level on the same
+			// page — never across a level change (real nesting) or once body
+			// text has begun.
+			if current.level > 0 && newLevel == current.level &&
+				current.title != "" &&
+				strings.TrimSpace(current.body.String()) == "" &&
+				(current.pageEnd == 0 || row.page == 0 || row.page == current.pageEnd) {
+				current.title = strings.TrimSpace(current.title + " " + text)
+				touch(current, row.page)
+				continue
+			}
+			current = &flat{level: newLevel, title: text}
 			touch(current, row.page)
 			flats = append(flats, current)
 			continue
@@ -509,6 +531,11 @@ func (p *PDF) parseDoc(_ context.Context, buf []byte) (*ParsedDoc, error) {
 		}}
 	}
 
+	// Drop empty "structural" leaf fragments (multi-line title remnants,
+	// running headers) so no section returns empty content to the browser /
+	// API; their titles ride onto the sibling body they head.
+	rootSec.Children = foldEmptyLeafSections(rootSec.Children)
+
 	// Internal sections inherit the union of their children's page ranges
 	// so callers reading the outline can still cite a page span.
 	propagateSectionPages(rootSec.Children)
@@ -550,6 +577,115 @@ func (p *PDF) resolvedParseTimeout() time.Duration {
 	return p.ParseTimeout
 }
 
+// foldEmptyLeafSections eliminates leaf sections (no children) whose
+// content is empty. A PDF whose heading detector fires on a fragment that
+// carries no body — a lone display-title line the row-merge didn't catch,
+// a repeated running header, a stray bold word — otherwise ships as a
+// "purely structural" Section that returns empty content to the section
+// browser and the /v1/sections/{id} API, which reads as a broken/empty
+// part of the document.
+//
+// The fold is lossless: an empty leaf's TITLE is preserved by attaching it
+// as a bold markdown heading line to the content of the sibling it heads —
+// the NEXT sibling when one exists (the body it introduces), else the
+// PREVIOUS sibling. Consecutive empty leaves accumulate and flush together
+// into the first sibling that can carry them, so a multi-line title that
+// escaped the row-merge still lands as one heading block above its body.
+// Page ranges union so citations stay correct. Internal nodes (sections
+// WITH children) are recursed into but never dropped — an empty heading
+// with real sub-sections is legitimate structure whose content lives in
+// its children.
+//
+// A lone empty leaf with no siblings to carry its title is kept as-is
+// (dropping it would erase the only trace of that heading); this is the
+// rare degenerate case and harms nothing.
+func foldEmptyLeafSections(sections []Section) []Section {
+	for i := range sections {
+		if len(sections[i].Children) > 0 {
+			sections[i].Children = foldEmptyLeafSections(sections[i].Children)
+		}
+	}
+
+	// Collect the indices of empty leaves eligible to fold. A section is an
+	// empty leaf when it has no children and no non-whitespace content.
+	emptyLeaf := func(s *Section) bool {
+		return len(s.Children) == 0 && strings.TrimSpace(s.Content) == ""
+	}
+	// Nothing to do unless there is at least one empty leaf AND at least one
+	// sibling that can carry a folded title.
+	empties, carriers := 0, 0
+	for i := range sections {
+		if emptyLeaf(&sections[i]) {
+			empties++
+		} else {
+			carriers++
+		}
+	}
+	if empties == 0 || carriers == 0 {
+		return sections
+	}
+
+	out := make([]Section, 0, len(sections))
+	var pendingTitles []string
+	var pendStart, pendEnd int
+	takePending := func(target *Section) {
+		if len(pendingTitles) == 0 {
+			return
+		}
+		var b strings.Builder
+		for _, t := range pendingTitles {
+			b.WriteString("**")
+			b.WriteString(t)
+			b.WriteString("**\n\n")
+		}
+		b.WriteString(target.Content)
+		target.Content = strings.TrimSpace(b.String())
+		target.PageStart = minNonZero(pendStart, target.PageStart)
+		if pendEnd > target.PageEnd {
+			target.PageEnd = pendEnd
+		}
+		pendingTitles = nil
+		pendStart, pendEnd = 0, 0
+	}
+
+	for i := range sections {
+		s := sections[i]
+		if emptyLeaf(&s) {
+			if t := strings.TrimSpace(s.Title); t != "" {
+				pendingTitles = append(pendingTitles, t)
+				pendStart = minNonZero(pendStart, s.PageStart)
+				if s.PageEnd > pendEnd {
+					pendEnd = s.PageEnd
+				}
+			}
+			continue // drop the empty leaf; its title rides on pendingTitles
+		}
+		// A real section: it carries any pending titles as heading prefix
+		// (these are the fragments it follows).
+		takePending(&s)
+		out = append(out, s)
+	}
+
+	// Any titles still pending had no following carrier — attach them to the
+	// last emitted sibling as a trailing heading block so nothing is lost.
+	if len(pendingTitles) > 0 && len(out) > 0 {
+		last := &out[len(out)-1]
+		var b strings.Builder
+		b.WriteString(last.Content)
+		for _, t := range pendingTitles {
+			b.WriteString("\n\n**")
+			b.WriteString(t)
+			b.WriteString("**")
+		}
+		last.Content = strings.TrimSpace(b.String())
+		if pendEnd > last.PageEnd {
+			last.PageEnd = pendEnd
+		}
+		last.PageStart = minNonZero(pendStart, last.PageStart)
+	}
+	return out
+}
+
 // propagateSectionPages fills internal-node PageStart/PageEnd from the union
 // of descendant leaf ranges where the internal node didn't have its own
 // (because its body was empty / hoisted into children). Leaves keep their
@@ -1268,7 +1404,9 @@ func parsePDFWithOutline(outline pdflib.Outline, rows []pdfRow) (*ParsedDoc, boo
 		title = rootSec.Children[0].Title
 	}
 
-	// Propagate page ranges so internal nodes span their children.
+	// Fold empty leaf fragments, then propagate page ranges so internal
+	// nodes span their children.
+	rootSec.Children = foldEmptyLeafSections(rootSec.Children)
 	propagateSectionPages(rootSec.Children)
 
 	return &ParsedDoc{

pkg/retrieval/treewalk_strategy.go

@@ -140,15 +140,23 @@ const defaultTreeWalkMaxHops = 8
 // excerpt. Matches TreeWalk's reference behaviour.
 const defaultPageContentLimit = 16000
 
-// defaultTreeWalkMaxCitations bounds the FINAL cited-range set. Three
-// is generous for the answer-spans-one-place common case (where ONE is
-// ideal) while still allowing a genuinely multi-location answer (e.g. a
-// 10-K figure cross-referenced between the income statement and a
-// footnote) to cite two or three distinct ranges. The FinanceBench
-// signal that motivated the cap: confident single-pick = f1 1.0,
-// 5-range spray = f1 0. Capping at 3 keeps the legitimate multi-range
-// case while removing the long tail of low-confidence noise.
-const defaultTreeWalkMaxCitations = 3
+// defaultTreeWalkMaxCitations bounds the FINAL cited-range set. The cap
+// is a CEILING, not a target: it only bites when the model emits MORE
+// distinct ranges than this. The prompt still drives the model to the
+// minimal sufficient set (usually one), so a confident single-pick answer
+// still yields ONE citation and the FinanceBench precision behaviour is
+// unchanged — the signal that motivated a cap at all (confident single =
+// f1 1.0, 5-range spray = f1 0) is enforced by the prompt + confidence
+// gating, not by clamping legitimate multi-location answers.
+//
+// It is raised from 3 to 6 so richer, genuinely multi-location answers —
+// a synthesis question over a whitepaper that draws on the intro, a
+// method section and a results table; a comparison spanning several parts
+// of a report — surface every location they actually rely on rather than
+// being truncated to three. Six is still well below the "spray" regime the
+// prompt/confidence machinery suppresses, so precision-critical single-pick
+// answers are untouched while multi-source answers become fully traceable.
+const defaultTreeWalkMaxCitations = 6
 
 // strategyNameTreeWalk is the stable identifier for config
 // (retrieval.strategy: treewalk) and telemetry.