Encapsulate stream queue context and consolidate forwarding wrappers

- Create struct stream_emit_ctx packing stream queue arguments into 1 cache line
- Reduce emit_stream_bulk_wrapped from 13 parameters to 3
- Reduce flush_stream_queue from 10 parameters to 1
- Extract stream_queue_finish to eliminate duplicated cleanup in kernel instantiations
- Replace EMIT_STREAM_ROW macros with typed inline functions
- Move pending_wrap clearing directly into term_screen_erase, insert, delete, and scroll
- Remove sc_erase_display, sc_erase_line, sc_insert_chars, and sc_delete_chars wrappers
- Consolidate term_row and remove duplicated row_fill_gap helper

6 files changed, 249 insertions, 268 deletions

diff --git a/src/core.c b/src/core.c
index 5d50db9..fcd73d0 100644
--- a/src/core.c
+++ b/src/core.c
@@ -91,7 +91,7 @@ void term_sc_print_glyph(struct term *t, uint32_t cp, int width, uint16_t flags)
     int right = t->lr_margin_mode ? t->right_margin : cols - 1;
 
     if (TERM_EXPECT(width == 2 && cols > 1 && cx == right && t->autowrap, 0)) {
-        row_fill_gap(t, p, row, cx);
+        term_row_fill_gap(srow, cx);
         clear_wide_edges(row, cx, right + 1, srow->len);
         row[cx] = (struct cell){ ' ', TERM_COLOR_FG, TERM_COLOR_BG, CELL_WRAP_PAD, 0 };
         srow->len = (uint16_t)cols;
@@ -113,7 +113,7 @@ void term_sc_print_glyph(struct term *t, uint32_t cp, int width, uint16_t flags)
     } else {
         flags &= ~CELL_WIDE;
     }
-    row_fill_gap(t, p, row, cx);
+    term_row_fill_gap(srow, cx);
     if (TERM_EXPECT(t->insert_mode, 0)) {
         if (cx == srow->len)
             row[cx] = (struct cell){ ' ', TERM_COLOR_FG, TERM_COLOR_BG, 0, 0 };
@@ -333,13 +333,11 @@ void term_sc_tab(struct term *t, int n)
 
 void term_sc_scroll_region(struct term *t, int top, int bottom, int n, bool down)
 {
-    t->pending_wrap = false;
     term_screen_scroll_region(t, top, bottom, n, down, false);
 }
 
 void term_sc_scroll(struct term *t, int n, bool down)
 {
-    t->pending_wrap = false;
     term_screen_scroll_region(t, t->scroll_top, t->scroll_bottom, n, down, true);
 }
 
@@ -587,67 +585,100 @@ static void flush_orig_rows(struct term *t, struct screen *s,
     }
 }
 
-static void flush_stream_queue(struct term *t, struct screen *s,
-                               struct hist_ascii_line *q, int *q_count, int q_cap,
-                               int *orig_y, int *orig_rem, size_t *flushed_scrolled,
-                               const struct pen_attr *attr, bool def_pen)
+struct stream_emit_ctx {
+    struct term *term;
+    struct screen *screen;
+    struct hist_ascii_line *q;
+    const struct pen_attr *attr;
+    size_t flushed_scrolled;
+    int q_cap;
+    int q_count;
+    int orig_y;
+    int orig_rem;
+    int cols;
+    bool def_pen;
+};
+
+static void flush_stream_queue(struct stream_emit_ctx *ctx)
 {
-    if (*orig_rem > 0)
-        flush_orig_rows(t, s, orig_y, orig_rem, *orig_rem);
-    int flush_n = *q_count - q_cap;
-    (void)def_pen;
-    term_hist_push_ascii_lines(&t->hist, q, flush_n, attr);
-    if (q_cap > 0) {
-        if (flush_n >= q_cap)
-            memcpy(q, q + flush_n, (size_t)q_cap * sizeof(q[0]));
+    if (ctx->orig_rem > 0)
+        flush_orig_rows(ctx->term, ctx->screen, &ctx->orig_y, &ctx->orig_rem, ctx->orig_rem);
+    int flush_n = ctx->q_count - ctx->q_cap;
+    (void)ctx->def_pen;
+    term_hist_push_ascii_lines(&ctx->term->hist, ctx->q, flush_n, ctx->attr);
+    if (ctx->q_cap > 0) {
+        if (flush_n >= ctx->q_cap)
+            memcpy(ctx->q, ctx->q + flush_n, (size_t)ctx->q_cap * sizeof(ctx->q[0]));
         else
-            memmove(q, q + flush_n, (size_t)q_cap * sizeof(q[0]));
+            memmove(ctx->q, ctx->q + flush_n, (size_t)ctx->q_cap * sizeof(ctx->q[0]));
     }
-    *q_count = q_cap;
-    *flushed_scrolled += (size_t)flush_n;
+    ctx->q_count = ctx->q_cap;
+    ctx->flushed_scrolled += (size_t)flush_n;
 }
 
-static inline void emit_stream_bulk_wrapped(struct term *t, struct screen *s,
-                                            const char *src, int bulk_lines, int cols,
-                                            const struct pen_attr *attr, bool def_pen,
-                                            struct hist_ascii_line *q, int q_cap,
-                                            int *q_count, int *orig_y, int *orig_rem,
-                                            size_t *flushed_scrolled)
+static inline void emit_stream_bulk_wrapped(struct stream_emit_ctx *ctx,
+                                            const char *src, int bulk_lines)
 {
+    struct term *t = ctx->term;
+    struct screen *s = ctx->screen;
+    int cols = ctx->cols;
     if (TERM_EXPECT(t->pending_wrap, 0)) {
         row_set_wrapped(screen_get_row(s, s->rows - 1), true);
         t->pending_wrap = false;
         t->cursor_x = 0;
         term_screen_scroll_up(t);
     }
-    if (bulk_lines < q_cap) {
+    if (bulk_lines < ctx->q_cap) {
         const char *p = src;
         for (int r = 0; r < bulk_lines; r++, p += cols) {
-            if (*q_count == 256) {
-                flush_stream_queue(t, s, q, q_count, q_cap, orig_y, orig_rem,
-                                   flushed_scrolled, attr, def_pen);
+            if (ctx->q_count == 256) {
+                flush_stream_queue(ctx);
             }
-            q[(*q_count)++] = (struct hist_ascii_line){ p, (uint16_t)cols, 1 };
+            ctx->q[ctx->q_count++] = (struct hist_ascii_line){ p, (uint16_t)cols, 1 };
         }
         return;
     }
 
-    if (*orig_rem > 0)
-        flush_orig_rows(t, s, orig_y, orig_rem, *orig_rem);
-    if (*q_count > 0) {
-        term_hist_push_ascii_lines(&t->hist, q, *q_count, attr);
-        *flushed_scrolled += (size_t)*q_count;
-        *q_count = 0;
+    if (ctx->orig_rem > 0)
+        flush_orig_rows(t, s, &ctx->orig_y, &ctx->orig_rem, ctx->orig_rem);
+    if (ctx->q_count > 0) {
+        term_hist_push_ascii_lines(&t->hist, ctx->q, ctx->q_count, ctx->attr);
+        ctx->flushed_scrolled += (size_t)ctx->q_count;
+        ctx->q_count = 0;
     }
-    int hist_rows = bulk_lines - q_cap;
+    int hist_rows = bulk_lines - ctx->q_cap;
     if (hist_rows > 0) {
-        term_hist_push_ascii(&t->hist, src, hist_rows * cols, attr, true);
-        *flushed_scrolled += (size_t)hist_rows;
+        term_hist_push_ascii(&t->hist, src, hist_rows * cols, ctx->attr, true);
+        ctx->flushed_scrolled += (size_t)hist_rows;
     }
     const char *sp = src + (size_t)hist_rows * (size_t)cols;
-    for (int r = 0; r < q_cap; r++, sp += cols)
-        q[r] = (struct hist_ascii_line){ sp, (uint16_t)cols, 1 };
-    *q_count = q_cap;
+    for (int r = 0; r < ctx->q_cap; r++, sp += cols)
+        ctx->q[r] = (struct hist_ascii_line){ sp, (uint16_t)cols, 1 };
+    ctx->q_count = ctx->q_cap;
+}
+
+static inline TERM_ALWAYS_INLINE void emit_stream_row_fast(struct stream_emit_ctx *ctx,
+                                                           const char *row_ptr,
+                                                           int row_len, bool row_wrapped)
+{
+    if (TERM_EXPECT(ctx->q_count == 256, 0))
+        flush_stream_queue(ctx);
+    ctx->q[ctx->q_count++] = (struct hist_ascii_line){
+        row_ptr, (uint16_t)row_len, (uint8_t)row_wrapped
+    };
+}
+
+static inline TERM_ALWAYS_INLINE void emit_stream_row(struct stream_emit_ctx *ctx,
+                                                      const char *row_ptr,
+                                                      int row_len, bool row_wrapped)
+{
+    if (TERM_EXPECT(ctx->term->pending_wrap, 0)) {
+        row_set_wrapped(screen_get_row(ctx->screen, ctx->screen->rows - 1), true);
+        ctx->term->pending_wrap = false;
+        ctx->term->cursor_x = 0;
+        term_screen_scroll_up(ctx->term);
+    }
+    emit_stream_row_fast(ctx, row_ptr, row_len, row_wrapped);
 }
 
 static TERM_NOINLINE size_t stream_ascii_wrapped(struct term *t, const char *buf, size_t len)
@@ -877,6 +908,140 @@ static inline uint64_t line_masks(const char *buf, uint64_t *nl)
     return controls;
 }
 
+static void stream_queue_finish(struct stream_emit_ctx *ctx)
+{
+    int cols = ctx->cols;
+    size_t queued_rows = 0;
+    for (int j = 0; j < ctx->q_count; j++)
+        queued_rows += ctx->q[j].wrapped ? 1 : rows_narrow(ctx->q[j].len, cols);
+    size_t loop_scrolled = ctx->flushed_scrolled + queued_rows;
+    if (loop_scrolled == 0)
+        return;
+
+    struct term *t = ctx->term;
+    struct screen *s = ctx->screen;
+    int rows = s->rows;
+
+    int push_orig = queued_rows < (size_t)ctx->orig_rem ? (int)queued_rows : ctx->orig_rem;
+    if (push_orig > 0)
+        flush_orig_rows(t, s, &ctx->orig_y, &ctx->orig_rem, push_orig);
+
+    int needed = rows - 1 - ctx->orig_rem;
+    int tail_rows = 0;
+    int k = ctx->q_count;
+    while (k > 0 && tail_rows < needed) {
+        k--;
+        struct hist_ascii_line pr = ctx->q[k];
+        int r = (pr.wrapped || pr.len == 0) ? 1 : (int)rows_narrow(pr.len, cols);
+        tail_rows += r;
+    }
+
+    int flush_n = k;
+    if (flush_n > 0) {
+        term_hist_push_ascii_lines(&t->hist, ctx->q, flush_n, ctx->attr);
+    }
+
+    int skip_rows = tail_rows - needed;
+    if (skip_rows > 0) {
+        struct hist_ascii_line pr = ctx->q[k];
+        int skip_bytes = skip_rows * cols;
+        if (skip_bytes > (int)pr.len) skip_bytes = (int)pr.len;
+        term_hist_push_ascii(&t->hist, pr.ptr, skip_bytes, ctx->attr, true);
+    }
+
+    if (ctx->orig_rem == 0) {
+        s->top = 0;
+    } else {
+        s->top += tail_rows;
+        if (s->top >= rows)
+            s->top -= rows;
+    }
+    /* Rows that stay on the screen become text rows. A row that
+     * continues the row above starts from its pen. */
+    int cur_y = ctx->orig_rem;
+    bool prev_wrapped = false;
+    for (int j = k; j < ctx->q_count && cur_y < rows - 1; j++) {
+        struct hist_ascii_line pr = ctx->q[j];
+        int p_len = pr.len;
+        const char *p_ptr = pr.ptr;
+
+        if (p_len == 0) {
+            int p = phys_row(t, cur_y);
+            struct srow *dst = &s->slot[p];
+            dst->uid = ++t->next_row_uid;
+            dst->version++;
+            uint8_t tail_padded = (dst->kind == ROW_CELLS && dst->len == 0)
+                                  ? (dst->flags & RF_TAIL_PADDED) : 0;
+            dst->kind = ROW_CELLS;
+            dst->len = 0;
+            dst->flags = tail_padded;
+            prev_wrapped = false;
+            cur_y++;
+            continue;
+        }
+
+        int off_in_line = 0;
+        if (j == k && skip_rows > 0) {
+            off_in_line = skip_rows * cols;
+            prev_wrapped = true;
+        }
+
+        while (off_in_line < p_len && cur_y < rows - 1) {
+            int chunk = p_len - off_in_line;
+            bool row_wraps = false;
+            if (chunk > cols) {
+                chunk = cols;
+                row_wraps = true;
+            } else if (pr.wrapped && off_in_line + chunk == p_len) {
+                row_wraps = true;
+            }
+
+            int p = phys_row(t, cur_y);
+            struct srow *dst = &s->slot[p];
+            dst->uid = ++t->next_row_uid;
+            dst->version++;
+
+            struct pen_attr start = prev_wrapped ? *ctx->attr : pen_attr_default();
+            dst->kind = ROW_TEXT;
+            dst->len = 0;
+            dst->flags = row_wraps ? RF_WRAPPED : 0;
+            text_row_start(&dst->text, &start);
+            dst->text.text = dst->ascii;
+            text_append(dst, (const uint8_t *)(p_ptr + off_in_line),
+                        chunk, chunk, chunk, HLINE_NARROW,
+                        ctx->attr->fg, ctx->attr->bg, pen_sl(ctx->attr));
+
+            prev_wrapped = row_wraps;
+            off_in_line += chunk;
+            cur_y++;
+        }
+    }
+    while (cur_y < rows - 1) {
+        int p = phys_row(t, cur_y);
+        struct srow *dst = &s->slot[p];
+        dst->uid = ++t->next_row_uid;
+        dst->version++;
+        dst->kind = ROW_CELLS;
+        dst->len = 0;
+        dst->flags = 0;
+        cur_y++;
+    }
+    int bot_p = phys_row(t, rows - 1);
+    struct srow *bot = &s->slot[bot_p];
+    uint8_t bot_tail = (bot->kind == ROW_CELLS && bot->len == 0)
+                       ? (bot->flags & RF_TAIL_PADDED) : 0;
+    bot->uid = ++t->next_row_uid;
+    bot->version++;
+    bot->kind = ROW_CELLS;
+    bot->len = 0;
+    bot->flags = bot_tail;
+    if (t->prompt_y > 0) {
+        t->prompt_y = (size_t)t->prompt_y > loop_scrolled
+                      ? t->prompt_y - (int)loop_scrolled : 0;
+    }
+    clamp_selection_after_scroll(t);
+}
+
 /* One body for every level. isa is a constant in each copy below, so each
  * copy keeps only its own kernels. */
 static inline TERM_ALWAYS_INLINE size_t bulk_ascii_lines_body(struct term *t, const char *buf,
@@ -894,7 +1059,6 @@ static inline TERM_ALWAYS_INLINE size_t bulk_ascii_lines_body(struct term *t, co
     bool autowrap = t->autowrap;
     bool lnm = term_lnm(t);
     int last_char = (int)t->last_char;
-    size_t flushed_scrolled = 0;
 
     struct pen_attr attr = {
         .fg = t->cur_fg,
@@ -909,29 +1073,19 @@ static inline TERM_ALWAYS_INLINE size_t bulk_ascii_lines_body(struct term *t, co
         return 0;
 
     struct hist_ascii_line q[256];
-    int q_cap = rows - 1;
-    int q_count = 0;
-    int orig_y = 0;
-    int orig_rem = q_cap;
-
-#define EMIT_STREAM_ROW_FAST(row_ptr, row_len, row_wrapped) do { \
-    if (TERM_EXPECT(q_count == 256, 0)) \
-        flush_stream_queue(t, s, q, &q_count, q_cap, &orig_y, &orig_rem, \
-                           &flushed_scrolled, &attr, def_pen); \
-    q[q_count++] = (struct hist_ascii_line){ \
-        (row_ptr), (uint16_t)(row_len), (uint8_t)(row_wrapped) \
-    }; \
-} while (0)
-
-#define EMIT_STREAM_ROW(row_ptr, row_len, row_wrapped) do { \
-    if (TERM_EXPECT(t->pending_wrap, 0)) { \
-        row_set_wrapped(screen_get_row(s, rows - 1), true); \
-        t->pending_wrap = false; \
-        t->cursor_x = 0; \
-        term_screen_scroll_up(t); \
-    } \
-    EMIT_STREAM_ROW_FAST(row_ptr, row_len, row_wrapped); \
-} while (0)
+    struct stream_emit_ctx ctx = {
+        .term = t,
+        .screen = s,
+        .q = q,
+        .attr = &attr,
+        .flushed_scrolled = 0,
+        .q_cap = rows - 1,
+        .q_count = 0,
+        .orig_y = 0,
+        .orig_rem = rows - 1,
+        .cols = cols,
+        .def_pen = def_pen,
+    };
 
     while (i < len) {
         /* Scan each block once. Keep the unfinished line in the input buffer. */
@@ -953,7 +1107,7 @@ static inline TERM_ALWAYS_INLINE size_t bulk_ascii_lines_body(struct term *t, co
                         goto narrow_done;
                     if (end > i)
                         last_char = (unsigned char)buf[end - 1];
-                    EMIT_STREAM_ROW_FAST(buf + i, (int)(end - i), false);
+                    emit_stream_row_fast(&ctx, buf + i, (int)(end - i), false);
                     i = end + 1;
                     nl &= nl - 1;
                 }
@@ -979,9 +1133,7 @@ narrow_done:
 
                     int bulk_lines = run / cols - (run % cols == 0);
                     if (bulk_lines > 0 && autowrap && (i + run == len || ((unsigned char)buf[i + run] >= 0x20 && (unsigned char)buf[i + run] < 0x7f) || (run % cols == 0))) {
-                        emit_stream_bulk_wrapped(t, s, buf + i, bulk_lines, cols, &attr, def_pen,
-                                                 q, q_cap, &q_count, &orig_y, &orig_rem,
-                                                 &flushed_scrolled);
+                        emit_stream_bulk_wrapped(&ctx, buf + i, bulk_lines);
                         size_t write_cells = (size_t)bulk_lines * (size_t)cols;
                         last_char = (unsigned char)buf[i + write_cells - 1];
                         i += write_cells;
@@ -998,7 +1150,7 @@ narrow_done:
 
                     if (eol > 0) {
                         last_char = (unsigned char)buf[i + limit - 1];
-                        EMIT_STREAM_ROW_FAST(buf + i, limit, false);
+                        emit_stream_row_fast(&ctx, buf + i, limit, false);
                         i += (size_t)(limit + eol);
                         continue;
                     }
@@ -1008,7 +1160,7 @@ narrow_done:
                                                    (unsigned char)buf[i + limit - 1]);
                     if (!wrapped) break;
                     last_char = (unsigned char)buf[i + limit - 1];
-                    EMIT_STREAM_ROW_FAST(buf + i, limit, true);
+                    emit_stream_row_fast(&ctx, buf + i, limit, true);
                     i += (size_t)limit;
                     continue;
                 }
@@ -1030,7 +1182,7 @@ narrow_done:
 
                 if (p1 > 0)
                     last_char = (unsigned char)buf[i + p1 - 1];
-                EMIT_STREAM_ROW_FAST(buf + i, p1, false);
+                emit_stream_row_fast(&ctx, buf + i, p1, false);
 
                 int next_start = p1 + eol1;
                 int last_consumed = next_start;
@@ -1051,7 +1203,7 @@ narrow_done:
                             if (eol2 > 0) {
                                 if (rel_p2 > 0)
                                     last_char = (unsigned char)buf[i + p2 - 1];
-                                EMIT_STREAM_ROW_FAST(buf + i + next_start, rel_p2, false);
+                                emit_stream_row_fast(&ctx, buf + i + next_start, rel_p2, false);
 
                                 int next_start2 = p2 + eol2;
                                 last_consumed = next_start2;
@@ -1072,7 +1224,7 @@ narrow_done:
                                             if (eol3 > 0) {
                                                 if (rel_p3 > 0)
                                                     last_char = (unsigned char)buf[i + p3 - 1];
-                                                EMIT_STREAM_ROW_FAST(buf + i + next_start2, rel_p3, false);
+                                                emit_stream_row_fast(&ctx, buf + i + next_start2, rel_p3, false);
 
                                                 last_consumed = p3 + eol3;
                                             }
@@ -1101,7 +1253,7 @@ narrow_done:
         if (TERM_EXPECT(eol > 0, 1)) {
             if (run > 0)
                 last_char = (unsigned char)buf[i + run - 1];
-            EMIT_STREAM_ROW(buf + i, run, false);
+            emit_stream_row(&ctx, buf + i, run, false);
             i += (size_t)(run + eol);
             continue;
         }
@@ -1109,7 +1261,7 @@ narrow_done:
         if (!run) {
             int neol = newline_bytes(buf + i, len - i);
             if (neol == 2 || (neol == 1 && lnm)) {
-                EMIT_STREAM_ROW(buf + i, 0, false);
+                emit_stream_row(&ctx, buf + i, 0, false);
                 i += (size_t)neol;
                 continue;
             }
@@ -1128,150 +1280,23 @@ narrow_done:
              ((unsigned char)buf[i + full] >= 0x20 &&
               (unsigned char)buf[i + full] < 0x7f) ||
              full % cols == 0)) {
-            emit_stream_bulk_wrapped(t, s, buf + i, bulk_lines, cols, &attr, def_pen,
-                                     q, q_cap, &q_count, &orig_y, &orig_rem,
-                                     &flushed_scrolled);
+            emit_stream_bulk_wrapped(&ctx, buf + i, bulk_lines);
             size_t write_cells = (size_t)bulk_lines * (size_t)cols;
             last_char = (unsigned char)buf[i + write_cells - 1];
             i += write_cells;
             continue;
         }
         last_char = (unsigned char)buf[i + run - 1];
-        EMIT_STREAM_ROW(buf + i, run, true);
+        emit_stream_row(&ctx, buf + i, run, true);
         i += (size_t)run;
     }
-#undef EMIT_STREAM_ROW
-#undef EMIT_STREAM_ROW_FAST
 #if defined(TERM_KERNEL_AVX512)
 done:
 #endif
     if (i == 0)
         return 0;
 
-    size_t queued_rows = 0;
-    for (int j = 0; j < q_count; j++)
-        queued_rows += q[j].wrapped ? 1 : rows_narrow(q[j].len, cols);
-    size_t loop_scrolled = flushed_scrolled + queued_rows;
-    if (loop_scrolled > 0) {
-        int push_orig = queued_rows < (size_t)orig_rem ? (int)queued_rows : orig_rem;
-        if (push_orig > 0)
-            flush_orig_rows(t, s, &orig_y, &orig_rem, push_orig);
-
-        int needed = rows - 1 - orig_rem;
-        int tail_rows = 0;
-        int k = q_count;
-        while (k > 0 && tail_rows < needed) {
-            k--;
-            struct hist_ascii_line pr = q[k];
-            int r = (pr.wrapped || pr.len == 0) ? 1 : (int)rows_narrow(pr.len, cols);
-            tail_rows += r;
-        }
-
-        int flush_n = k;
-        if (flush_n > 0) {
-            term_hist_push_ascii_lines(&t->hist, q, flush_n, &attr);
-        }
-
-        int skip_rows = tail_rows - needed;
-        if (skip_rows > 0) {
-            struct hist_ascii_line pr = q[k];
-            int skip_bytes = skip_rows * cols;
-            if (skip_bytes > (int)pr.len) skip_bytes = (int)pr.len;
-            term_hist_push_ascii(&t->hist, pr.ptr, skip_bytes, &attr, true);
-        }
-
-        if (orig_rem == 0) {
-            s->top = 0;
-        } else {
-            s->top += tail_rows;
-            if (s->top >= rows)
-                s->top -= rows;
-        }
-        /* Rows that stay on the screen become text rows. A row that
-         * continues the row above starts from its pen. */
-        int cur_y = orig_rem;
-        bool prev_wrapped = false;
-        for (int j = k; j < q_count && cur_y < rows - 1; j++) {
-            struct hist_ascii_line pr = q[j];
-            int p_len = pr.len;
-            const char *p_ptr = pr.ptr;
-
-            if (p_len == 0) {
-                int p = phys_row(t, cur_y);
-                struct srow *dst = &s->slot[p];
-                dst->uid = ++t->next_row_uid;
-                dst->version++;
-                uint8_t tail_padded = (dst->kind == ROW_CELLS && dst->len == 0)
-                                      ? (dst->flags & RF_TAIL_PADDED) : 0;
-                dst->kind = ROW_CELLS;
-                dst->len = 0;
-                dst->flags = tail_padded;
-                prev_wrapped = false;
-                cur_y++;
-                continue;
-            }
-
-            int off_in_line = 0;
-            if (j == k && skip_rows > 0) {
-                off_in_line = skip_rows * cols;
-                prev_wrapped = true;
-            }
-
-            while (off_in_line < p_len && cur_y < rows - 1) {
-                int chunk = p_len - off_in_line;
-                bool row_wraps = false;
-                if (chunk > cols) {
-                    chunk = cols;
-                    row_wraps = true;
-                } else if (pr.wrapped && off_in_line + chunk == p_len) {
-                    row_wraps = true;
-                }
-
-                int p = phys_row(t, cur_y);
-                struct srow *dst = &s->slot[p];
-                dst->uid = ++t->next_row_uid;
-                dst->version++;
-
-                struct pen_attr start = prev_wrapped ? attr : pen_attr_default();
-                dst->kind = ROW_TEXT;
-                dst->len = 0;
-                dst->flags = row_wraps ? RF_WRAPPED : 0;
-                text_row_start(&dst->text, &start);
-                dst->text.text = dst->ascii;
-                text_append(dst, (const uint8_t *)(p_ptr + off_in_line),
-                            chunk, chunk, chunk, HLINE_NARROW,
-                            attr.fg, attr.bg, pen_sl(&attr));
-
-                prev_wrapped = row_wraps;
-                off_in_line += chunk;
-                cur_y++;
-            }
-        }
-        while (cur_y < rows - 1) {
-            int p = phys_row(t, cur_y);
-            struct srow *dst = &s->slot[p];
-            dst->uid = ++t->next_row_uid;
-            dst->version++;
-            dst->kind = ROW_CELLS;
-            dst->len = 0;
-            dst->flags = 0;
-            cur_y++;
-        }
-        int bot_p = phys_row(t, rows - 1);
-        struct srow *bot = &s->slot[bot_p];
-        uint8_t bot_tail = (bot->kind == ROW_CELLS && bot->len == 0)
-                           ? (bot->flags & RF_TAIL_PADDED) : 0;
-        bot->uid = ++t->next_row_uid;
-        bot->version++;
-        bot->kind = ROW_CELLS;
-        bot->len = 0;
-        bot->flags = bot_tail;
-        if (t->prompt_y > 0) {
-            t->prompt_y = (size_t)t->prompt_y > loop_scrolled
-                          ? t->prompt_y - (int)loop_scrolled : 0;
-        }
-        clamp_selection_after_scroll(t);
-    }
+    stream_queue_finish(&ctx);
 
     t->last_char = (uint32_t)last_char;
     t->cursor_x = 0;
diff --git a/src/core.h b/src/core.h
index 2460938..beec4e0 100644
--- a/src/core.h
+++ b/src/core.h
@@ -103,34 +103,6 @@ static inline void sc_erase_span(struct term *t, int y, int from, int to)
     t->scr.slot[p].version++;
 }
 
-/* Erase display mode 0, 1, 2, or 3. */
-static inline void sc_erase_display(struct term *t, int mode)
-{
-    t->pending_wrap = false;
-    term_screen_erase(t, mode);
-}
-
-/* Erase line mode 0, 1, or 2. */
-static inline void sc_erase_line(struct term *t, int mode)
-{
-    t->pending_wrap = false;
-    term_screen_erase_line(t, mode);
-}
-
-/* Insert n blank characters at cursor. */
-static inline void sc_insert_chars(struct term *t, int n)
-{
-    t->pending_wrap = false;
-    term_screen_insert_chars(t, n);
-}
-
-/* Delete n characters at cursor. */
-static inline void sc_delete_chars(struct term *t, int n)
-{
-    t->pending_wrap = false;
-    term_screen_delete_chars(t, n);
-}
-
 /* Scroll margins: top to bottom, n lines up or down. */
 void term_sc_scroll_region(struct term *t, int top, int bottom, int n, bool down);
 
diff --git a/src/escape.c b/src/escape.c
index 9576e8e..fc254af 100644
--- a/src/escape.c
+++ b/src/escape.c
@@ -746,7 +746,7 @@ static inline TERM_ALWAYS_INLINE void dispatch_csi(struct term *t, const struct
         }
         break;
     case 'J':
-        sc_erase_display(t, csi->params[0]);
+        term_screen_erase(t, csi->params[0]);
         break;
     case 'X': {
         int n = csi_count(csi);
@@ -759,13 +759,13 @@ static inline TERM_ALWAYS_INLINE void dispatch_csi(struct term *t, const struct
             t->decsace = (uint8_t)csi->params[0];
         break;
     case 'K':
-        sc_erase_line(t, csi->params[0]);
+        term_screen_erase_line(t, csi->params[0]);
         break;
     case 'P':
-        sc_delete_chars(t, csi_count(csi));
+        term_screen_delete_chars(t, csi_count(csi));
         break;
     case '@':
-        sc_insert_chars(t, csi_count(csi));
+        term_screen_insert_chars(t, csi_count(csi));
         break;
     case 'c': {
         if (csi->prefix == '>')
diff --git a/src/fastcsi.c b/src/fastcsi.c
index d8983dc..74d4800 100644
--- a/src/fastcsi.c
+++ b/src/fastcsi.c
@@ -169,13 +169,13 @@ int term_try_fast_csi(struct term *t, const char *buf, size_t rem)
                    (p[0] >= '0' && p[0] <= '3' && p + 1 < end && p[1] == 'J')) {
             t->cur_fg = fg;
             t->cur_bg = bg;
-            sc_erase_display(t, p[0] == 'J' ? 0 : p[0] - '0');
+            term_screen_erase(t, p[0] == 'J' ? 0 : p[0] - '0');
             p += p[0] == 'J' ? 1 : 2;
         } else if (p[0] == 'K' ||
                    (p[0] >= '0' && p[0] <= '2' && p + 1 < end && p[1] == 'K')) {
             t->cur_fg = fg;
             t->cur_bg = bg;
-            sc_erase_line(t, p[0] == 'K' ? 0 : p[0] - '0');
+            term_screen_erase_line(t, p[0] == 'K' ? 0 : p[0] - '0');
             p += p[0] == 'K' ? 1 : 2;
         } else if (!t->origin_mode && !t->lr_margin_mode &&
                    ((*p >= '0' && *p <= '9') || *p == ';' ||
@@ -279,13 +279,13 @@ int term_try_fast_csi(struct term *t, const char *buf, size_t rem)
             case 'J':
                 t->cur_fg = fg;
                 t->cur_bg = bg;
-                sc_erase_display(t, pv[0]);
+                term_screen_erase(t, pv[0]);
                 p = q + 1;
                 break;
             case 'K':
                 t->cur_fg = fg;
                 t->cur_bg = bg;
-                sc_erase_line(t, pv[0]);
+                term_screen_erase_line(t, pv[0]);
                 p = q + 1;
                 break;
             case 'X':
@@ -298,13 +298,13 @@ int term_try_fast_csi(struct term *t, const char *buf, size_t rem)
             case 'P':
                 t->cur_fg = fg;
                 t->cur_bg = bg;
-                sc_delete_chars(t, n);
+                term_screen_delete_chars(t, n);
                 p = q + 1;
                 break;
             case '@':
                 t->cur_fg = fg;
                 t->cur_bg = bg;
-                sc_insert_chars(t, n);
+                term_screen_insert_chars(t, n);
                 p = q + 1;
                 break;
             case 's':
diff --git a/src/screen.c b/src/screen.c
index afd9b4a..d0f573c 100644
--- a/src/screen.c
+++ b/src/screen.c
@@ -324,9 +324,9 @@ void term_screen_erase_line(struct term *t, int mode)
 
 void term_screen_delete_chars(struct term *t, int n)
 {
+    t->pending_wrap = false;
     if (t->cursor_x >= t->cols)
         return;
-    t->pending_wrap = false;
     int x = t->cursor_x;
     int cols = t->cols;
     if (n > cols - x)
@@ -365,9 +365,9 @@ void term_screen_delete_chars(struct term *t, int n)
 
 void term_screen_insert_chars(struct term *t, int n)
 {
+    t->pending_wrap = false;
     if (t->cursor_x >= t->cols)
         return;
-    t->pending_wrap = false;
     int x = t->cursor_x;
     int cols = t->cols;
     if (n > cols - x)
@@ -407,6 +407,7 @@ void term_screen_insert_chars(struct term *t, int n)
 
 void term_screen_erase_all(struct term *t)
 {
+    t->pending_wrap = false;
     struct screen *s = &t->scr;
     bool pristine = using_default_colors(t);
     for (int y = 0; y < s->rows; y++) {
@@ -544,6 +545,7 @@ static inline bool srow_has_text(const struct srow *r, int cols)
 
 void term_screen_erase(struct term *t, int mode)
 {
+    t->pending_wrap = false;
     if (mode == 3) {
         term_hist_clear(&t->hist);
         term_hist_view_clear(&t->view);
@@ -667,6 +669,7 @@ static void rotate_and_clear_srows(struct term *t, struct srow *base, int height
 void term_screen_scroll_region(struct term *t, int top, int bottom, int count,
                                bool down, bool history)
 {
+    t->pending_wrap = false;
     if (count <= 0 || top > bottom || top < 0 || bottom >= t->rows)
         return;
     int height = bottom - top + 1;
diff --git a/src/screen.h b/src/screen.h
index 50e18c6..8498e9d 100644
--- a/src/screen.h
+++ b/src/screen.h
@@ -161,15 +161,7 @@ static inline struct srow *screen_row_at(const struct screen *s, int y)
 
 static inline struct srow *term_row(const struct term *t, int y)
 {
-    int p = t->scr.top + y;
-    struct srow *row = &t->scr.slot[p >= t->scr.rows ? p - t->scr.rows : p];
-    srow_materialize_text(row);
-    return row;
-}
-
-static inline struct srow *term_screen_row(const struct term *t, int y)
-{
-    return term_row(t, y);
+    return screen_get_row(&t->scr, y);
 }
 
 /* Write the cells of a text row to out: its text, then its fill blanks.
@@ -361,17 +353,6 @@ static inline void mark_row_blank(struct term *t, int p, uint32_t fg, uint32_t b
     (void)bg;
 }
 
-static inline void row_fill_gap(struct term *t, int p, struct cell *row, int x)
-{
-    struct screen *s = term_active_screen(t);
-    struct srow *r = &s->slot[p];
-    if (x > r->len) {
-        if (!(r->flags & RF_TAIL_PADDED))
-            fill_cells(row + r->len, x - r->len, TERM_COLOR_FG, TERM_COLOR_BG);
-        r->len = (uint16_t)x;
-    }
-}
-
 static inline void term_row_fill_gap(struct srow *row, int x)
 {
     if (x > row->len) {
@@ -413,7 +394,7 @@ static inline void erase_span(struct term *t, int p, int from, int to)
         row_set_wrapped(row, false);
 
     if (!pristine) {
-        row_fill_gap(t, p, cells, from);
+        term_row_fill_gap(row, from);
         fill_cells(cells + from, to - from, t->cur_fg, t->cur_bg);
         if (to > len) {
             row->len = (uint16_t)to;