fix(ab-corpus): forbid mixing batch_design with element tools in T

The earlier T instruction split tool selection into PRIMARY / COMPOSITE
(element tools) and FALLBACK (batch_design). FALLBACK was scoped
"when no element tool fits a given component shape" — implying you
could mix per-component. But output-parser.ts silently drops every
batch_design tag whenever any element call is also present (filter
ELEMENT_TOOL_NAME_RE then return). A mixed response

  <op_tool>{"name":"add_section_header_v0",...}</op_tool>
  <op_tool>{"name":"batch_design", ...scaffolding...}</op_tool>

would lose the batch_design half and only run the element call —
the brief is half-applied without anyone noticing. Codex stop-time
review caught the mismatch.

Reframes T as a binary choice:
- STRATEGY A: every component fits an add_*_v0 tool — emit one tag
  per component
- STRATEGY B: at least one component needs batch_design — emit a
  SINGLE batch_design covering the whole brief

Plus an explicit "Do not mix Strategy A and Strategy B" guard line.
Tests assert both strategy markers are present in every T variant
and the dropped per-component-fallback phrase is gone.

The elements.md cookbook still teaches batch_design + element tool
composition for real MCP multi-round usage; the corpus T arm is
single-shot so the trailing T_TOOL_CALL_INSTRUCTIONS overrides via
"last instruction wins".

scripts/ab-corpus/__tests__/build-prompt.test.ts

@@ -9,16 +9,21 @@ import { buildSystemPrompt } from '../build-prompt';
 // the diet silently no-ops.
 
 const T_MULTI_TOOL_MARKER = 'Respond with one or more `<op_tool>` tags';
-const T_COMPOSITE_MARKER = 'COMPOSITE (multi-component briefs';
+const T_STRATEGY_A_MARKER = 'STRATEGY A — element tools';
+const T_STRATEGY_B_MARKER = 'STRATEGY B — batch_design fallback';
+const T_NO_MIX_MARKER = 'Do not mix Strategy A and Strategy B';
 const B_BATCH_MARKER =
   '<op_tool>{"name": "batch_design", "arguments": {"operations": "<DSL_STRING>"}}</op_tool>';
 // Earlier versions of the T instructions explicitly forbade multi-tool
 // output ("Respond with one tag" / "Do not combine multiple tags");
 // that wording undercut the multi-tool teaching baked into elements.md.
-// Codex stop-time review caught the contradiction — guard against
-// regression by asserting the forbidding text is GONE.
+// A subsequent revision flipped to allow chaining but quietly invited
+// mixed batch_design + element tags per-component, which the corpus
+// parser silently drops. Codex stop-time review caught both. Guard
+// against re-introducing either failure mode.
 const FORBIDDEN_SINGLE_TAG_PHRASE = 'one `<op_tool>` tag, nothing else';
 const FORBIDDEN_NO_COMBINE_PHRASE = 'Do not combine multiple tags';
+const FORBIDDEN_PER_COMPONENT_FALLBACK_PHRASE = 'when no element tool fits a given component shape';
 
 describe('buildSystemPrompt', () => {
   it('T + composite returns the largest prompt (cookbook present)', () => {
@@ -64,11 +69,12 @@ describe('buildSystemPrompt', () => {
     expect(distinct.size).toBe(1);
   });
 
-  it('every T variant carries the multi-tool output-format marker', () => {
+  it('every T variant carries the multi-tool + Strategy A/B output-format markers', () => {
     for (const difficulty of ['obvious', 'optional', 'composite'] as const) {
       const built = buildSystemPrompt('T', { difficulty });
       expect(built.system).toContain(T_MULTI_TOOL_MARKER);
-      expect(built.system).toContain(T_COMPOSITE_MARKER);
+      expect(built.system).toContain(T_STRATEGY_A_MARKER);
+      expect(built.system).toContain(T_STRATEGY_B_MARKER);
     }
   });
 
@@ -80,9 +86,21 @@ describe('buildSystemPrompt', () => {
     }
   });
 
-  it('B variant carries the batch_design marker but not the multi-tool split', () => {
+  it('T variant explicitly forbids mixing Strategy A and Strategy B (parser drops mixed)', () => {
+    for (const difficulty of ['obvious', 'optional', 'composite'] as const) {
+      const built = buildSystemPrompt('T', { difficulty });
+      expect(built.system).toContain(T_NO_MIX_MARKER);
+      // Earlier wording invited per-component fallback ("when no element
+      // tool fits a given component shape") which the parser silently
+      // drops when element calls are also present. Guard the rephrase.
+      expect(built.system).not.toContain(FORBIDDEN_PER_COMPONENT_FALLBACK_PHRASE);
+    }
+  });
+
+  it('B variant carries the batch_design marker but not the Strategy A/B split', () => {
     const built = buildSystemPrompt('B');
     expect(built.system).toContain(B_BATCH_MARKER);
-    expect(built.system).not.toContain(T_COMPOSITE_MARKER);
+    expect(built.system).not.toContain(T_STRATEGY_A_MARKER);
+    expect(built.system).not.toContain(T_STRATEGY_B_MARKER);
   });
 });

scripts/ab-corpus/build-prompt.ts

@@ -45,21 +45,18 @@ const T_TOOL_CALL_INSTRUCTIONS = [
   '',
   'OUTPUT FORMAT — EMIT AS TOOL CALL(S):',
   '',
-  'Respond with one or more `<op_tool>` tags, nothing else. The harness reads every `<op_tool>` tag in your output, so chain as many as the brief implies. Choose tools based on intent:',
+  'Respond with one or more `<op_tool>` tags, nothing else. The harness reads every `<op_tool>` tag in your output, so chain as many as the brief implies. Pick ONE strategy for the whole response — do NOT mix element tools with batch_design in the same output:',
   '',
-  'PRIMARY (single component): emit one tag for the add_*_v0 element tool that fits.',
-  '  <op_tool>{"name": "add_X_v0", "arguments": {...}}</op_tool>',
-  '',
-  'COMPOSITE (multi-component briefs — settings panel with N rows, team list with N members, audit feed with N entries, onboarding screen with N steps): emit ONE `<op_tool>` per component under the same parent. Tags appear in render order (top-to-bottom for vertical, left-to-right for horizontal).',
+  'STRATEGY A — element tools (preferred): when every component in the brief fits an `add_*_v0` element tool, emit one tag per component, in render order (top-to-bottom for vertical layouts, left-to-right for horizontal). Single-component briefs produce exactly one tag; multi-component briefs (settings panel with N rows, team list with N members, audit feed with N entries, onboarding screen with N step cards) produce N+1 or more.',
   '  <op_tool>{"name": "add_section_header_v0", "arguments": {"title": "Notifications"}}</op_tool>',
   '  <op_tool>{"name": "add_setting_row_v0", "arguments": {...}}</op_tool>',
   '  <op_tool>{"name": "add_setting_row_v0", "arguments": {...}}</op_tool>',
   '',
-  'FALLBACK: when no element tool fits a given component shape, emit a single batch_design call.',
+  'STRATEGY B — batch_design fallback: when the brief includes any component shape that NO element tool covers (heterogeneous custom layout, post-hoc styling, bespoke scaffolding), emit a SINGLE batch_design call covering the WHOLE response. Multiple `<op_tool>` tags carrying batch_design or any mix of batch_design + element tools is not supported — the harness drops the batch_design half and only the element calls run.',
   '  <op_tool>{"name": "batch_design", "arguments": {"operations": "<DSL_STRING>"}}</op_tool>',
-  'The `operations` value is a single string containing the batch_design DSL.',
+  'The `operations` value is a single string containing the batch_design DSL covering every component in the brief.',
   '',
-  'Do not add prose before, after, or between tags.',
+  'Do not add prose before, after, or between tags. Do not mix Strategy A and Strategy B in the same output — choose element tools for every component or batch_design for every component.',
   '',
 ].join('\n');