Fix Compare size labels and experiment explainer copy

dashboard/experiments.css

@@ -49,6 +49,11 @@ body.dash-view-experiments #filter-policy-meta {
   margin: 0.5rem 0 0;
 }
 
+.exp-why[open] {
+  max-height: min(28rem, 55vh);
+  overflow: auto;
+}
+
 .exp-story p {
   margin: 0.35rem 0;
 }

dashboard/experiments.js

@@ -244,43 +244,44 @@ function selectHtml(id, label, values, current, labelFn) {
 
 function renderWhyExperiments() {
   return `
-    <details class="exp-howto exp-why glass-panel" open>
+    <details class="exp-howto exp-why glass-panel">
       <summary>Why experiments, if the Dashboard already has every number?</summary>
       <p>
-        The main Dashboard is the <strong>full warehouse</strong>: every library, every record shape.
-        That is the right place to browse. It is the wrong place to pick a library for a real service
-        by sorting on speed.
+        The main Dashboard is the <strong>full warehouse</strong>: every library, every payload size,
+        every pattern. That is the right place to browse. It is the wrong place to pick a queue
+        for a real service by sorting on speed.
       </p>
       <p>
-        The fastest row in the big table may be a library you <strong>cannot use</strong>.
-        It might write bytes that only one language can read.
-        It might drop field names so partners cannot parse the text.
-        It might be built for a file you write once, not for a web request you write on every click.
-        The Dashboard will still put it first, because it is fast.
+        The fastest row in the big table may be a queue you <strong>cannot use</strong>.
+        It might only work with one writer and one reader. It might be an async queue
+        while your workers are operating-system threads. It might be a scheduler that
+        does not hand a payload from a producer to a consumer. The Dashboard will still
+        put it first, because it is fast on that slice.
       </p>
       <p>
         An experiment starts with a <strong>decision</strong>, not a sort.
-        Example: “We must keep JSON on the public website. Which JSON library is fast enough?”
-        We keep only the libraries that could be the answer. We time one shared record.
+        Example: “One request thread puts work; one worker takes it. The ticket is
+        256 bytes. Which in-process queue is fast enough?”
+        We keep only the libraries that could be the answer. We time one shared payload.
         We write down what we would give up.
       </p>
       <p>
-        Think of a grocery store and a recipe. The store has every product — that is the main Dashboard.
-        The recipe lists only what you need for dinner, and why — that is an experiment.
-        You need both. The store does not tell you what to cook.
+        Think of a grocery store and a recipe. The store has every product — that is the
+        main Dashboard. The recipe lists only what you need for dinner, and why — that
+        is an experiment. You need both. The store does not tell you what to cook.
       </p>
       <p>
         Concrete traps the big table does not prevent:
       </p>
       <ul>
-        <li><strong>Different jobs in one list.</strong> JSON, compact bytes, and a Python-only format can share a table. “What is fastest?” is not one question.</li>
-        <li><strong>A tiny gap looks like a winner.</strong> 1.7 µs vs 1.8 µs is not a reason to rewrite a service.</li>
-        <li><strong>Your constraint is invisible.</strong> If browsers must keep JSON, the big table still shows every format.</li>
-        <li><strong>One record is not one hundred.</strong> The winner at 1 can lose at 100.</li>
-        <li><strong>Numbers without a story.</strong> “Binary is faster” is how teams break a public website.</li>
+        <li><strong>Different jobs in one list.</strong> A 1P1C ring and a 4P4C channel can share a table. “What is fastest?” is not one question.</li>
+        <li><strong>A tiny gap looks like a winner.</strong> A few percent is not a reason to rewrite a service.</li>
+        <li><strong>Your constraint is invisible.</strong> If you must use threads, the big table still shows async queues.</li>
+        <li><strong>One size is not another.</strong> The winner at 256 bytes can lose at 4 KiB.</li>
+        <li><strong>One pattern is not another.</strong> The winner at 1P1C can lose at 4P4C.</li>
       </ul>
       <p>
-        We do not run a second benchmark. We cut the same numbers down to a fair contest
+        We do not run a second clock. We cut the same numbers down to a fair contest
         and put the story next to them.
         <a href="../experiments/">Longer notes</a>
       </p>

dashboard/main.js

@@ -1878,7 +1878,7 @@ function populateDataTypeSelect(options, cfg = {}) {
     batchCompound.forEach(addOpt);
   }
   if (allTypes.length) {
-    addSep('── compounded data types ──');
+    addSep('── compounded data sizes ──');
     allTypes.forEach(addOpt);
   }
   if (allAll.length) {
@@ -2231,20 +2231,20 @@ function discoverDataTypeOptions(allGroups) {
   }
   batchCompound.sort();
 
-  // Cross-type at fixed n
+  // Both published sizes at one item count: all@100, all@1000.
   const allTypes = [];
-  if (nsGlobal.has(1)) allTypes.push('all@1');
-  if (nsGlobal.has(100)) allTypes.push('all@100');
-
-  // Everything
-  const allAll = natural.length ? ['all@all'] : [];
+  if (byBase.size >= 2) {
+    for (const n of [...nsGlobal].sort((a, b) => a - b)) {
+      allTypes.push(`all@${n}`);
+    }
+  }
 
   return {
     natural,
     batchCompound,
     allTypes,
-    allAll,
-    all: [...natural, ...batchCompound, ...allTypes, ...allAll],
+    allAll: [],
+    all: [...natural, ...batchCompound, ...allTypes],
   };
 }
 
@@ -3420,7 +3420,7 @@ function renderTable() {
       scopeNote =
         ` <strong>Compounded batch</strong>: mean of <code>${escapeHtml(scope.base)}@n=${scope.nA}</code> and <code>${escapeHtml(scope.base)}@n=${scope.nB}</code>.`;
     } else if (scope.kind === 'all_n') {
-      scopeNote = ` <strong>Compounded data types</strong> at <code>n=${scope.n}</code>.`;
+      scopeNote = ` <strong>Both sizes</strong> at <code>${scope.n}</code> items.`;
     } else if (scope.kind === 'all_all') {
       scopeNote = ` <strong>Compounded all</strong> (<code>all@all</code>).`;
     }
@@ -3730,7 +3730,7 @@ function copyRosterMarkdown() {
         return `Scope: compounded batch ${state.currentTestData} · mode ${state.currentMode}`;
       }
       if (s.kind === 'all_n') {
-        return `Scope: all data types @ n=${s.n} · mode ${state.currentMode}`;
+        return `Scope: both sizes · ${s.n} items · ${modeDisplayLabel(normalizeMode(state.currentMode))}`;
       }
       if (s.kind === 'all_all') {
         return `Scope: all@all (all types × all n) · mode ${state.currentMode}`;

dashboard/public/data/experiments/index.json

@@ -1,6 +1,6 @@
 {
   "schema": "gld.dashboard.experiments/1",
-  "generated_at": "2026-09-01T22:23:00Z",
+  "generated_at": "2026-09-02T00:00:42Z",
   "count": 13,
   "with_results": 13,
   "experiments": [
@@ -598,9 +598,9 @@
       "number": 13,
       "slug": "payload-size-sweep",
       "title": "Which payload sizes change the ranking?",
-      "question": "Same SPSC handoff, five payload sizes (1 B, 64 B, 256 B, 4 KiB, 64 KiB). Which sizes produce a different ranking, and which are just more of the same?\n",
+      "question": "Same 1P1C handoff, five payload sizes (1 B, 64 B, 256 B, 4 KiB, 64 KiB). Which sizes produce a different ranking, and which are just more of the same?\n",
       "story": {
-        "why": "The default suite names five data types, but every runner builds an opaque byte string of a chosen length. If ranking only moves at a couple of lengths, the published matrix should keep those lengths and drop the rest.\n",
+        "why": "Every runner builds an opaque byte string of a chosen length. If ranking only moves at a couple of lengths, the published matrix should keep those lengths and drop the rest.\n",
         "example": "A ticket is a few hundred bytes. An HTTP body is a few kilobytes. A file chunk can be tens of kilobytes. We need to know which of those worlds actually changes which queue wins.\n",
         "tradeoff": "Extra sizes cost machine time and make the dashboard harder to read. One size is enough if copy cost never changes the order. Two sizes are enough if a small payload and a large payload tell different stories.\n"
       },

experiments/13-payload-size-sweep/README.md

@@ -1,16 +1,14 @@
 # 13 — Payload size sweep
 
-**Question:** which payload sizes change the SPSC ranking, and which are
+**Question:** which payload sizes change the 1P1C ranking, and which are
 redundant?
 
-This is not another named data type. Every runner already builds an opaque
-byte string. The five catalog names (`message`, `event`, `telemetry`,
-`strings`, `document`) are five lengths between 256 B and 4 KiB. This
-folder asks whether the published suite needs those names at all, or only
-one or two lengths.
+Every runner already builds an opaque byte string of a chosen length.
+This folder asked which lengths change ranking, and which are just more
+of the same.
 
 Sizes in this run: **1 B, 64 B, 256 B, 4 KiB, 64 KiB**. One hundred items
-per repetition. SPSC only. One hundred repetitions (`full`).
+per repetition. 1P1C only. One hundred repetitions (`full`).
 
 ```bash
 ./experiments/13-payload-size-sweep/run.sh python

experiments/13-payload-size-sweep/experiment.yaml

@@ -2,15 +2,14 @@ schema: gld.experiment.config/1
 id: 13-payload-size-sweep
 title: Which payload sizes change the ranking?
 question: >
-  Same SPSC handoff, five payload sizes (1 B, 64 B, 256 B, 4 KiB, 64 KiB).
+  Same 1P1C handoff, five payload sizes (1 B, 64 B, 256 B, 4 KiB, 64 KiB).
   Which sizes produce a different ranking, and which are just more of the
   same?
 story:
   why: >
-    The default suite names five data types, but every runner builds an
-    opaque byte string of a chosen length. If ranking only moves at a
-    couple of lengths, the published matrix should keep those lengths and
-    drop the rest.
+    Every runner builds an opaque byte string of a chosen length. If ranking
+    only moves at a couple of lengths, the published matrix should keep those
+    lengths and drop the rest.
   example: >
     A ticket is a few hundred bytes. An HTTP body is a few kilobytes. A
     file chunk can be tens of kilobytes. We need to know which of those

experiments/13-payload-size-sweep/results.json

@@ -1,6 +1,6 @@
 {
   "generated_at": "2026-09-01T22:23:00Z",
-  "question": "Which SPSC payload sizes change the ranking?",
+  "question": "Which 1P1C payload sizes change the ranking?",
   "sizes_bytes": [
     1,
     64,

experiments/13-payload-size-sweep/results.md

@@ -2,7 +2,7 @@
 
 Times are not comparable across languages.
 
-Question: which SPSC payload sizes change the ranking?
+Question: which 1P1C payload sizes change the ranking?
 
 ## Recommendation
 
@@ -24,9 +24,9 @@ Do not confuse **payload bytes per item** with **how many items** move in one re
 - `size_1` (1 B): same first place as 256 B, no copy-bound pack collapse, no inversion.
 - `size_64` (64 B): same first place as 256 B, no copy-bound pack collapse, no inversion.
 
-## What this means for the named types
+## What this means for the published matrix
 
-The published catalog names `message`, `event`, `telemetry`, `strings`, and `document` are already five lengths: 256 B, 512 B, 1 KiB, 2 KiB, and 4 KiB. Runners do not serialize those object graphs. The warehouse SPSC ranks on those names match this sweep: Python and C never change first place; JavaScript and C# only shuffle a mid-pack pair; Rust first becomes copy-bound at 4 KiB. 512 B, 1 KiB, and 2 KiB do not add a ranking question. Keep `message` (256 B) and `document` (4 KiB). Drop the three names in between from the default matrix when that change is applied.
+Every runner builds an opaque byte string. This sweep asked which lengths change 1P1C ranking. Python and C never change first place. JavaScript and C# only shuffle a close pair. Rust first changes at 4 KiB, when the queue copies every byte. Keep 256 B and 4 KiB on the default matrix. 1 B and 64 B match 256 B. 64 KiB is a later research cell, not a third published size.
 
 ## Per language vs 256 B
 

experiments/13-payload-size-sweep/summarize.py

@@ -371,7 +371,7 @@ def write_combined_md(
         "",
         "Times are not comparable across languages.",
         "",
-        "Question: which SPSC payload sizes change the ranking?",
+        "Question: which 1P1C payload sizes change the ranking?",
         "",
         "## Recommendation",
         "",
@@ -408,17 +408,14 @@ def write_combined_md(
         lines.append("")
     lines.extend(
         [
-            "## What this means for the named types",
+            "## What this means for the published matrix",
             "",
-            "The published catalog names `message`, `event`, `telemetry`, "
-            "`strings`, and `document` are already five lengths: 256 B, 512 B, "
-            "1 KiB, 2 KiB, and 4 KiB. Runners do not serialize those object "
-            "graphs. The warehouse SPSC ranks on those names match this sweep: "
-            "Python and C never change first place; JavaScript and C# only "
-            "shuffle a mid-pack pair; Rust first becomes copy-bound at 4 KiB. "
-            "512 B, 1 KiB, and 2 KiB do not add a ranking question. Keep "
-            "`message` (256 B) and `document` (4 KiB). Drop the three names "
-            "in between from the default matrix when that change is applied.",
+            "Every runner builds an opaque byte string. This sweep asked which "
+            "lengths change 1P1C ranking. Python and C never change first "
+            "place. JavaScript and C# only shuffle a close pair. Rust first "
+            "changes at 4 KiB, when the queue copies every byte. Keep 256 B "
+            "and 4 KiB on the default matrix. 1 B and 64 B match 256 B. "
+            "64 KiB is a later research cell, not a third published size.",
             "",
             "## Per language vs 256 B",
             "",
@@ -470,7 +467,7 @@ def main() -> int:
     recommendation = recommend(lang_reports)
     payload = {
         "generated_at": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
-        "question": "Which SPSC payload sizes change the ranking?",
+        "question": "Which 1P1C payload sizes change the ranking?",
         "sizes_bytes": [SIZE_BYTES[t] for t in SIZE_ORDER],
         "baseline_bytes": SIZE_BYTES[BASELINE],
         "recommendation": recommendation,