tests: a capture is a delta and a click names its word, so a tagline edit stops rewriting the suite

1 files changed, 223 insertions, 60 deletions

diff --git a/test/snapshot.zig b/test/snapshot.zig
index f2d7984c..7f8c4f21 100644
--- a/test/snapshot.zig
+++ b/test/snapshot.zig
@@ -48,6 +48,16 @@
 //   resize <rows> <cols>
 //   snap <label>              capture grid text + cursor
 //   snapstyle <label>         capture per-cell style runs (fg,bg,attrs)
+//
+// Any <col> above may be `@Word`, `@Word#2` or `@Word+4` instead of a number:
+// the column of a word on that row of the live grid (see resolveCol). Chrome
+// words move, and a frozen column silently clicks whatever moved into it.
+//
+// A capture is a DELTA against the previous capture of the same kind in the
+// same script: the first one (and any after a resize) is the whole screen as
+// `|<row>`, the rest are only the rows that changed, as `|<n>: <row>`. Before
+// this, every frame restated every row, so 82% of golden lines were a copy of
+// the line above and one added builtin word rewrote 78 goldens (see Prev).
 const std = @import("std");
 const libc = std.c;
 const ghostty_vt = @import("ghostty-vt");
@@ -105,13 +115,14 @@ pub fn main(init: std.process.Init) !void {
             script_count += 1;
         }
     }
-    // Updating is the one run whose output becomes evidence instead of being
-    // checked against it. The ordinary parallel pass can reject a transient
-    // capture and retry it alone; update mode has no expected bytes to compare,
-    // so it cannot know which capture needs that retry. Run it serially even if
-    // --jobs was supplied: a slower intentional baseline refresh is cheaper
-    // than blessing scheduler or shared-process startup noise as behavior.
-    if (update) jobs = 1;
+    // Updating used to be serial, because update mode has no expected bytes and
+    // so cannot tell a transient capture from behaviour. Serial was a proxy for
+    // that, and an expensive one (77s against 6s). What it was reaching for is
+    // available directly: capture at the WIDEST probe settings the harness has
+    // (beCareful, as if every script had already failed once), then run the
+    // ordinary verify pass over what was just written. A golden that does not
+    // reproduce is reported instead of blessed, and the parallel pass keeps its
+    // retry machinery, which serial update never had.
     var scripts = try arena.alloc([]const u8, script_count);
     var script_i: usize = 0;
     var saw_exe = false;
@@ -144,7 +155,13 @@ pub fn main(init: std.process.Init) !void {
     _ = lockOrWait(SNAP_BASE ++ "/.run.lock", "another pardes-snap run");
 
     const codes = try arena.alloc(u8, scripts.len);
-    try runBatch(arena, exe_z, toBatch(arena, scripts), update, jobs, codes);
+    if (update) {
+        careful = true;
+        try runBatch(arena, exe_z, toBatch(arena, scripts), true, jobs, codes);
+        careful = false;
+        std.debug.print("verifying {d} regenerated golden(s)\n", .{scripts.len});
+    }
+    try runBatch(arena, exe_z, toBatch(arena, scripts), false, jobs, codes);
 
     // A failure in the parallel pass is more often contention than regression:
     // the app can lose a race with a kworker while 20 siblings fight for the
@@ -179,7 +196,7 @@ pub fn main(init: std.process.Init) !void {
             std.debug.print("\n", .{});
             careful = round > 0;
             const recodes = try arena.alloc(u8, paths.len);
-            try runBatch(arena, exe_z, paths, update, 1, recodes);
+            try runBatch(arena, exe_z, paths, false, 1, recodes);
             for (retry, recodes) |i, c| codes[i] = c;
         }
     }
@@ -195,7 +212,11 @@ pub fn main(init: std.process.Init) !void {
         }
     }
     if (failed > 0) {
-        std.debug.print("{d}/{d} snapshot scripts FAILED\n", .{ failed, scripts.len });
+        // In update mode the golden was written and then failed to reproduce:
+        // the capture is not stable, so it is not a baseline.
+        std.debug.print("{d}/{d} snapshot scripts {s}\n", .{
+            failed, scripts.len, if (update) "UNSTABLE (regenerated, does not reproduce)" else "FAILED",
+        });
         std.process.exit(1);
     }
     std.debug.print("all {d} snapshot scripts ok\n", .{scripts.len});
@@ -382,6 +403,8 @@ fn runScript(arena: std.mem.Allocator, exe_z: [:0]const u8, script_path: []const
     var chunk_len: usize = 0;
     var h: ?eh.Harness = null;
     defer if (h) |*hp| hp.deinit();
+    // captures are deltas against the previous one of the same kind (see Prev)
+    var prev: Prev = .{};
 
     var lines_it = std.mem.splitScalar(u8, src, '\n');
     var lineno: usize = 0;
@@ -444,6 +467,9 @@ fn runScript(arena: std.mem.Allocator, exe_z: [:0]const u8, script_path: []const
                 extra[nextra] = (try arena.dupeZ(u8, a)).ptr;
             }
             if (h) |*hp| hp.deinit(); // scripts may restart (e.g. dump then -l)
+            // a new process is a new screen: capture it whole, or its first
+            // frame would be a delta against a dead instance's rows
+            prev = .{};
             h = try eh.Harness.initArgv(gpa, exe_z.ptr, rows, cols, extra[0..nextra]);
         } else if (std.mem.eql(u8, cmd, "snapfile")) {
             const label = tok.next() orelse return error.BadScript;
@@ -469,29 +495,32 @@ fn runScript(arena: std.mem.Allocator, exe_z: [:0]const u8, script_path: []const
             while (tok.next()) |name| try live(&h).send(keyBytes(name) orelse return error.BadKey);
         } else if (std.mem.eql(u8, cmd, "press") or std.mem.eql(u8, cmd, "release") or std.mem.eql(u8, cmd, "drag")) {
             const btn = buttonCode(tok.next() orelse return error.BadScript) orelse return error.BadScript;
-            const col = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col_spec = tok.next() orelse return error.BadScript;
             const row = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col = try resolveCol(live(&h), col_spec, row);
             const b: u16 = if (std.mem.eql(u8, cmd, "drag")) btn + 32 else btn;
             const fin: u8 = if (std.mem.eql(u8, cmd, "release")) 'm' else 'M';
             try live(&h).send(try std.fmt.allocPrint(arena, "\x1b[<{d};{d};{d}{c}", .{ b, col, row, fin }));
         } else if (std.mem.eql(u8, cmd, "motion")) {
-            const col = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col_spec = tok.next() orelse return error.BadScript;
             const row = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col = try resolveCol(live(&h), col_spec, row);
             try live(&h).send(try std.fmt.allocPrint(arena, "\x1b[<35;{d};{d}M", .{ col, row }));
         } else if (std.mem.eql(u8, cmd, "wheel")) {
             const dir = tok.next() orelse return error.BadScript;
             const b: u16 = if (std.mem.eql(u8, dir, "up")) 64 else 65;
-            const col = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col_spec = tok.next() orelse return error.BadScript;
             const row = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
+            const col = try resolveCol(live(&h), col_spec, row);
             try live(&h).send(try std.fmt.allocPrint(arena, "\x1b[<{d};{d};{d}M", .{ b, col, row }));
         } else if (std.mem.eql(u8, cmd, "resize")) {
             const rows = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
             const cols = try std.fmt.parseInt(u16, tok.next() orelse return error.BadScript, 10);
             try live(&h).resize(rows, cols);
         } else if (std.mem.eql(u8, cmd, "snap")) {
-            try appendChunk(chunks, &chunk_len, try snapText(arena, live(&h), tok.rest()));
+            try appendChunk(chunks, &chunk_len, try snapText(arena, live(&h), tok.rest(), &prev));
         } else if (std.mem.eql(u8, cmd, "snapstyle")) {
-            try appendChunk(chunks, &chunk_len, try snapStyle(arena, live(&h), tok.rest()));
+            try appendChunk(chunks, &chunk_len, try snapStyle(arena, live(&h), tok.rest(), &prev));
         } else {
             std.debug.print("unknown command {s}\n", .{cmd});
             return error.BadScript;
@@ -679,9 +708,20 @@ fn waitStable(h: *eh.Harness, quiet: i64, timeout: i64) !void {
 // mints a fresh pty). So the hold is a bet on kworker latency, and a loaded
 // machine loses it. The parallel pass bets small; the serial re-run that decides
 // the verdict bets big.
-var probe_after_ms: i64 = 10;
+//
+// Only the SUM is the safety margin: a return cannot happen before
+// `probe_after_ms + probe_hold_ms` after the last change either way. The split
+// is pure cost, and the two halves cost differently - the arm window is one
+// `poll()` sleep, while every `probe_every_ms` of hold is a full `treeAsleep()`
+// walk of the app's process tree. Measured across the suite: 796 armed probes,
+// 54 of them saw a later change, none later than 11ms at 20 jobs. That is the
+// scheduler-tick case. The number the sum is really sized for is bigger and
+// found by adversarial measurement: a theme transition sleeps
+// `animation.frame_ms` (16ms) between frames with the whole tree asleep, and
+// under `taskset -c 0,1` those gaps stretch to 57ms. 60ms total stays.
+var probe_after_ms: i64 = 50;
 const probe_every_ms: i64 = 5;
-var probe_hold_ms: i64 = 50;
+var probe_hold_ms: i64 = 10;
 
 /// Widen the probes for the re-run pass: nothing else is competing then, so the
 /// extra wait is cheap, and its verdict is the one that gets reported.
@@ -771,51 +811,179 @@ fn stateKey(h: *eh.Harness) ![]u8 {
     return std.fmt.allocPrint(h.gpa, "{d},{d},{x}\n{s}", .{ c.x, c.y, styles.final(), text });
 }
 
-fn snapText(arena: std.mem.Allocator, h: *eh.Harness, label: []const u8) ![]u8 {
+/// A mouse column: either a screen column, or `@Word` / `@Word#2` - the column
+/// of a WORD found on that row of the LIVE grid.
+///
+/// Chrome words move. When `Newtty` was added to every tail and `Joincol` and
+/// `Changelog` to the topbar, six scripts kept their frozen columns and started
+/// clicking a different word: `tutor.snap` clicked `Grep` instead of `Tutor`,
+/// `exec.snap`/`respawn.snap`/`tagalign.snap` clicked `Newtty` instead of `Del`,
+/// `find.snap` clicked `Joincol` instead of `Find`, `windowops.snap` clicked
+/// `Tutor` instead of `Debug` - and `--update` blessed all of it, so 256 golden
+/// lines were green while asserting the opposite of their script's first line.
+/// Naming the word instead of its column makes that class of drift impossible:
+/// the word is either there to be clicked or the script fails loudly.
+///
+/// The match is boundary-anchored, which is not a detail: `New` is a prefix of
+/// `Newcol` and of `Newtty`, and `@New` resolving to `Newcol` is exactly the bug
+/// this is here to prevent. `#n` picks the nth occurrence for a row that holds
+/// two panes' tags, and `+n`/`-n` addresses a column NEAR a word - which is how
+/// a click that means "in the gap the right-aligned builtins leave after the
+/// path" is spelled without depending on how wide those builtins are.
+fn resolveCol(h: *eh.Harness, spec: []const u8, row: u16) !u16 {
+    if (spec.len == 0 or spec[0] != '@') return std.fmt.parseInt(u16, spec, 10);
+    var word = spec[1..];
+    var want: usize = 1;
+    var offset: i32 = 0;
+    // parsed off the end, so a word may hold a `-` or a `#` itself
+    if (std.mem.lastIndexOfAny(u8, word, "+-")) |sign| {
+        if (sign > 0) {
+            if (std.fmt.parseInt(i32, word[sign + 1 ..], 10)) |n| {
+                offset = if (word[sign] == '-') -n else n;
+                word = word[0..sign];
+            } else |_| {}
+        }
+    }
+    if (std.mem.lastIndexOfScalar(u8, word, '#')) |hash| {
+        if (std.fmt.parseInt(usize, word[hash + 1 ..], 10)) |n| {
+            want = n;
+            word = word[0..hash];
+        } else |_| {}
+    }
+    if (word.len == 0 or want == 0) return error.BadScript;
+    if (row == 0 or row > h.rows) return error.WordRowMissing;
+    // One byte per CELL, not per byte of UTF-8: a mouse column is a cell, and
+    // a tag row can hold a wide glyph or the U+254E column divider left of the
+    // word. Non-ASCII cells become 0xff, which is not a word byte, so they
+    // read as boundaries - every chrome word is ASCII.
+    var cells: [1024]u8 = undefined;
+    const line = rowCells(h, row - 1, &cells);
+    var seen: usize = 0;
+    var from: usize = 0;
+    while (std.mem.indexOfPos(u8, line, from, word)) |at| : (from = at + 1) {
+        if (at > 0 and wordByte(line[at - 1])) continue;
+        const after = at + word.len;
+        if (after < line.len and wordByte(line[after])) continue;
+        seen += 1;
+        if (seen != want) continue;
+        // SGR columns are 1-based
+        const col = @as(i32, @intCast(at + 1)) + offset;
+        if (col < 1 or col > @as(i32, h.cols)) return error.BadScript;
+        return @intCast(col);
+    }
+    std.debug.print("  no \"{s}\" #{d} on row {d}: {s}\n", .{ word, want, row, std.mem.trimEnd(u8, line, " ") });
+    return error.WordNotOnRow;
+}
+
+fn rowCells(h: *eh.Harness, y: u16, buf: *[1024]u8) []const u8 {
+    const n = @min(h.cols, buf.len);
+    var x: u16 = 0;
+    while (x < n) : (x += 1) {
+        const ci = h.term.screens.active.pages.getCell(.{ .viewport = .{ .x = x, .y = y } });
+        buf[x] = blk: {
+            const cell = (ci orelse break :blk ' ').cell;
+            if (cell.content_tag != .codepoint and cell.content_tag != .codepoint_grapheme) break :blk ' ';
+            const cp = cell.content.codepoint.data;
+            break :blk if (cp == 0) ' ' else if (cp < 128) @intCast(cp) else 0xff;
+        };
+    }
+    return buf[0..n];
+}
+
+fn wordByte(c: u8) bool {
+    return std.ascii.isAlphanumeric(c) or c == '_';
+}
+
+/// A capture, kept so the NEXT capture of the same kind can be emitted as the
+/// rows that changed instead of the whole screen again. 81.7% of re-captured
+/// rows used to be byte-identical to the previous capture in the same script,
+/// which is why one word added to a tagline rewrote 78 goldens: every frame
+/// re-stated every row of chrome. The delta keeps the assertion (a row that
+/// moves still shows up, and a capture whose only change is the cursor is
+/// exactly the empty delta its script means) and drops the restatement.
+const Prev = struct {
+    snap: []const []const u8 = &.{},
+    style: []const []const u8 = &.{},
+    // per kind: one shared width let a `snapstyle` between a `resize` and the
+    // next `snap` re-arm the guard, and the post-resize frame then came out as
+    // a delta against rows of the old width - dropping every row whose text is
+    // width-invariant, which is every left-aligned chrome row there is.
+    snap_cols: u16 = 0,
+    style_cols: u16 = 0,
+};
+
+/// `== snap <label> grid=CxR cursor=X,Y` then either the whole screen as
+/// `|<row text>` (first capture of a script, or after a resize) or just the
+/// rows that changed as `|<n>: <row text>`. An indexed row is a patch against
+/// the previous capture; an unindexed one is a frame.
+fn snapText(arena: std.mem.Allocator, h: *eh.Harness, label: []const u8, prev: *Prev) ![]u8 {
     const text = try h.screenText();
     defer h.gpa.free(text);
     const c = h.cursor();
-    const header = try std.fmt.allocPrint(arena, "== snap {s} grid={d}x{d} cursor={d},{d}\n", .{ label, h.cols, h.rows, c.x, c.y });
-    var total = header.len;
+    const rows = try arena.alloc([]const u8, h.rows);
     var it = std.mem.splitScalar(u8, text, '\n');
-    var row: usize = 0;
-    while (row < h.rows) : (row += 1) {
-        const line = it.next() orelse "";
-        total += 1 + std.mem.trimEnd(u8, line, " ").len + 1;
-    }
-    const out = try arena.alloc(u8, total);
-    var pos: usize = 0;
-    copyBytes(out, &pos, header);
-    it = std.mem.splitScalar(u8, text, '\n');
-    row = 0;
-    while (row < h.rows) : (row += 1) {
-        const line = it.next() orelse "";
-        copyBytes(out, &pos, "|");
-        copyBytes(out, &pos, std.mem.trimEnd(u8, line, " "));
-        copyBytes(out, &pos, "\n");
+    for (rows) |*r| r.* = try arena.dupe(u8, std.mem.trimEnd(u8, it.next() orelse "", " "));
+
+    var out: std.ArrayList(u8) = .empty;
+    try out.appendSlice(arena, try std.fmt.allocPrint(arena, "== snap {s} grid={d}x{d} cursor={d},{d}\n", .{
+        label, h.cols, h.rows, c.x, c.y,
+    }));
+    if (prev.snap.len == rows.len and prev.snap_cols == h.cols) {
+        for (rows, prev.snap, 0..) |now, was, i| {
+            if (std.mem.eql(u8, now, was)) continue;
+            // no trailing space when a row goes blank: a whitespace-trimming
+            // editor or `git apply --whitespace=fix` would rewrite it and the
+            // golden would then fail like a regression
+            try out.appendSlice(arena, if (now.len == 0)
+                try std.fmt.allocPrint(arena, "|{d}:\n", .{i})
+            else
+                try std.fmt.allocPrint(arena, "|{d}: {s}\n", .{ i, now }));
+        }
+    } else {
+        // A full frame stops at the last row with anything on it: the blank
+        // tail of the grid is the same information as the `grid=CxR` header.
+        var end = rows.len;
+        while (end > 0 and rows[end - 1].len == 0) end -= 1;
+        for (rows[0..end]) |r| {
+            try out.appendSlice(arena, "|");
+            try out.appendSlice(arena, r);
+            try out.appendSlice(arena, "\n");
+        }
     }
-    return out;
+    prev.snap = rows;
+    prev.snap_cols = h.cols;
+    return out.items;
 }
 
 /// Per-row style runs from the harness emulator: `|y: x0-x1 fg,bg,attrs ...`.
 /// fg/bg: d (default) | pN (palette) | #rrggbb. attrs: b f i k r v x + u U ~ . -
-/// for underline single/double/curly/dotted/dashed.
-fn snapStyle(arena: std.mem.Allocator, h: *eh.Harness, label: []const u8) ![]u8 {
-    const header = try std.fmt.allocPrint(arena, "== style {s}\n", .{label});
-    const rows_len = styleRows(h, null);
-    const out = try arena.alloc(u8, header.len + rows_len);
-    var pos: usize = 0;
-    copyBytes(out, &pos, header);
-    _ = styleRows(h, out[pos..]);
-    return out;
+/// for underline single/double/curly/dotted/dashed. Deltas the same way
+/// `snapText` does - the row already carries its own index, so a patch and a
+/// frame are spelled identically here. The header carries the grid because a
+/// style capture has no cursor: without it an unchanged style would be a label
+/// line asserting nothing at all.
+fn snapStyle(arena: std.mem.Allocator, h: *eh.Harness, label: []const u8, prev: *Prev) ![]u8 {
+    const rows = try styleRows(arena, h);
+    var out: std.ArrayList(u8) = .empty;
+    try out.appendSlice(arena, try std.fmt.allocPrint(arena, "== style {s} grid={d}x{d}\n", .{ label, h.cols, h.rows }));
+    const delta = prev.style.len == rows.len and prev.style_cols == h.cols;
+    for (rows, 0..) |now, i| {
+        if (delta and std.mem.eql(u8, now, prev.style[i])) continue;
+        try out.appendSlice(arena, "|");
+        try out.appendSlice(arena, now);
+        try out.appendSlice(arena, "\n");
+    }
+    prev.style = rows;
+    prev.style_cols = h.cols;
+    return out.items;
 }
 
-fn styleRows(h: *eh.Harness, dst: ?[]u8) usize {
-    var pos: usize = 0;
-    var y: u16 = 0;
-    while (y < h.rows) : (y += 1) {
-        var fmt_buf: [64]u8 = undefined;
-        emitBytes(dst, &pos, std.fmt.bufPrint(&fmt_buf, "|{d}:", .{y}) catch unreachable);
+fn styleRows(arena: std.mem.Allocator, h: *eh.Harness) ![]const []const u8 {
+    const out = try arena.alloc([]const u8, h.rows);
+    for (out, 0..) |*dst, y_usize| {
+        const y: u16 = @intCast(y_usize);
+        var row: std.ArrayList(u8) = .empty;
+        try row.appendSlice(arena, try std.fmt.allocPrint(arena, "{d}:", .{y}));
         var key_buf: [64]u8 = undefined;
         var run_key: [64]u8 = undefined;
         var run_len: usize = 0;
@@ -829,20 +997,15 @@ fn styleRows(h: *eh.Harness, dst: ?[]u8) usize {
                 continue;
             }
             if (std.mem.eql(u8, key, run_key[0..run_len])) continue;
-            emitBytes(dst, &pos, std.fmt.bufPrint(&fmt_buf, " {d}-{d} ", .{ run_start, x - 1 }) catch unreachable);
-            emitBytes(dst, &pos, run_key[0..run_len]);
+            try row.appendSlice(arena, try std.fmt.allocPrint(arena, " {d}-{d} ", .{ run_start, x - 1 }));
+            try row.appendSlice(arena, run_key[0..run_len]);
             @memcpy(run_key[0..key.len], key);
             run_len = key.len;
             run_start = x;
         }
-        emitBytes(dst, &pos, "\n");
+        dst.* = row.items;
     }
-    return pos;
-}
-
-fn emitBytes(dst: ?[]u8, pos: *usize, bytes: []const u8) void {
-    if (dst) |out| @memcpy(out[pos.*..][0..bytes.len], bytes);
-    pos.* += bytes.len;
+    return out;
 }
 
 fn cellKey(h: *eh.Harness, x: u16, y: u16, buf: *[64]u8) []const u8 {