feat: prev button, step instructions all tabs, noise threshold 3.2, word corrections I->T G->s

smoke_signal/scripts/punct_corrector.py

@@ -26,7 +26,6 @@ SMART_QUOTES = {"“", "”", "‘", "’"}
 
 KNOWN_SUBSTITUTIONS: dict[str, str] = {
     "l": "!",
-    "I": "!",
     "|": "!",
     "1": "!",
     ",,": "\"",
@@ -35,6 +34,23 @@ KNOWN_SUBSTITUTIONS: dict[str, str] = {
     "ʼ": "'",
 }
 
+# Whole-word corrections for known Tesseract errors on Gruffalo-style fonts.
+# I misread as T at word start; G misread as s.
+# Extend this as more errors are found in review.
+WORD_CORRECTIONS: dict[str, str] = {
+    "Tndeed": "Indeed",
+    "Tll":    "I'll",
+    "Tt":     "It",
+    "Ti":     "It",
+    "sood":   "good",
+    "sreat":  "great",
+    "sot":    "got",
+    "sround": "ground",
+    "srew":   "grew",
+    "save":   "gave",
+    "soing":  "going",
+}
+
 MIN_PUNCT_PER_10_WORDS = 0.8
 
 
@@ -237,6 +253,15 @@ def apply_punctuation_corrections(
         result_chars.append(learned if learned else ch)
     text = "".join(result_chars)
 
+    # Apply whole-word corrections (Tesseract font-specific errors)
+    import re as _re
+    def _fix_word(m):
+        w = m.group(0)
+        return WORD_CORRECTIONS.get(w, w)
+    pattern = "|".join(_re.escape(k) for k in WORD_CORRECTIONS)
+    if pattern:
+        text = _re.sub(pattern, _fix_word, text)
+
     flags = validate_punctuation(text)
     words = _word_count(text)
     punct = _punct_count(text)

smoke_signal_tab.py

@@ -1752,7 +1752,7 @@ def _run_tesseract_batch(
                 printable = sum(1 for c in txt if c.isprintable() and not c.isspace())
                 r_alpha = (letters / printable) if printable else 0.0
                 # Keep if: long enough tokens OR high confidence OR single short word (punctuation line)
-                if r_avg_tok >= 3.0 or r["confidence"] >= 0.80 or (len(tokens) == 1 and r["confidence"] >= 0.60):
+                if r_avg_tok >= 3.2 or r["confidence"] >= 0.82 or (len(tokens) == 1 and r["confidence"] >= 0.65):
                     clean_regions.append(r)
             if not clean_regions:
                 # Everything filtered — page is pure illustration noise
@@ -2864,6 +2864,13 @@ def smoke_signal_tab():
                     <div class="ss-panel-icon">👁</div>
                     <div><p class="ss-panel-title">OCR Engine</p>
                     <p class="ss-panel-sub">Surya layout + recognition · Confidence scoring · Review queue</p></div>
+                </div>
+                <div style="background:#1a2535;border:1px solid #30363d;border-radius:8px;padding:12px 18px;margin:12px 0;font-size:12px;color:#8b949e;line-height:1.8">
+                    <b style="color:#e6edf3">&#9654; Order:</b>
+                    <span style="color:#3fb950">①</span> Ingest + Save Book Pages &nbsp;
+                    <span style="color:#3fb950">②</span> Click <b style="color:#f0883e">Run OCR</b> below &nbsp;
+                    <span style="color:#3fb950">③</span> Wait for log to complete &nbsp;
+                    <span style="color:#3fb950">④</span> Go to Review tab
                 </div>""")
 
                 ocr_status = gr.HTML(_ocr_status_html())
@@ -2932,6 +2939,13 @@ def smoke_signal_tab():
                     <div class="ss-panel-icon">✎</div>
                     <div><p class="ss-panel-title">Review Workbench</p>
                     <p class="ss-panel-sub">Correct · Accept · Reject · Build gold training set</p></div>
+                </div>
+                <div style="background:#1a2535;border:1px solid #30363d;border-radius:8px;padding:12px 18px;margin:12px 0;font-size:12px;color:#8b949e;line-height:1.8">
+                    <b style="color:#e6edf3">&#9654; Order:</b>
+                    <span style="color:#3fb950">①</span> Click <b style="color:#f0883e">↻ Load Review Queue</b> &nbsp;
+                    <span style="color:#3fb950">②</span> Review the page image and text &nbsp;
+                    <span style="color:#3fb950">③</span> Edit text if needed · click <b style="color:#f0883e">✓ Accept</b>, <b style="color:#f0883e">✎ Save Edit</b>, or <b style="color:#f0883e">✗ Reject</b> &nbsp;
+                    <span style="color:#3fb950">④</span> Use <b style="color:#f0883e">← Prev</b> / <b style="color:#f0883e">Next →</b> to navigate without deciding
                 </div>""")
 
                 review_status = gr.HTML(_review_status_html())
@@ -2972,7 +2986,9 @@ def smoke_signal_tab():
                         with gr.Row():
                             reject_btn = gr.Button("✗ Reject", elem_classes=["ss-btn-reject"])
                             quar_btn   = gr.Button("⚑ Quarantine", elem_classes=["ss-btn-quar"])
-                        next_btn = gr.Button("Next Item →", elem_classes=["ss-btn-next"])
+                        with gr.Row():
+                            prev_btn = gr.Button("← Prev", elem_classes=["ss-btn-next"])
+                            next_btn = gr.Button("Next →", elem_classes=["ss-btn-next"])
 
                 load_review_btn = gr.Button("↻ Load Review Queue", elem_classes=["ss-btn-next"])
                 current_idx = gr.State(0)
@@ -3034,6 +3050,13 @@ def smoke_signal_tab():
                     img, raw, info, _, _ = get_review_item(new_idx)
                     return new_idx, img, raw, raw, info
 
+                def _prev(idx):
+                    new_idx = max(0, idx - 1)
+                    img, raw, info, _, _ = get_review_item(new_idx)
+                    return new_idx, img, raw, raw, info
+
+                prev_btn.click(_prev, inputs=[current_idx],
+                               outputs=[current_idx, review_image, raw_text_box, final_text_box, item_info])
                 next_btn.click(_next, inputs=[current_idx],
                                outputs=[current_idx, review_image, raw_text_box, final_text_box, item_info])