From c8c30d5682e46c8db34973e6d2f94551782bddc2 Mon Sep 17 00:00:00 2001
From: Brummel <mould@mycelium.de>
Date: Wed, 20 May 2026 18:11:21 +0200
Subject: [PATCH] =?UTF-8?q?test(drift):=20RED=20=E2=80=94=20anchor-presenc?=
 =?UTF-8?q?e=20must=20scope=20to=20jsonc=20fenced=20blocks?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds `anchor_presence_check_is_scoped_to_jsonc_blocks` to
`design_schema_drift.rs`. The test pins the property in two
directions: an anchor mentioned only in prose must report ABSENT
under the scoped helper; an anchor inside a ```jsonc``` block must
report PRESENT; the live `design/contracts/data-model.md` must
remain PRESENT (anti-over-narrowing guard).

The helper `anchor_in_jsonc_block` does not yet exist — this test
compile-blocks the entire `design_schema_drift` file, which is the
intended contract pressure: the GREEN side must introduce the
helper for any drift test to run.

Pre-rolesplit the audit framed this as "anchors live in Decision 11
instead of §Data model"; the role-split iter (176821c) made
`data-model.md` its own file, but the per-match scope was never
narrowed — `.contains()` still cannot tell jsonc-block anchors
(load-bearing) from prose anchors (incidental). This RED pins the
remaining surface.

refs #10
---
 .../ailang-core/tests/design_schema_drift.rs  | 76 +++++++++++++++++++
 1 file changed, 76 insertions(+)

diff --git a/crates/ailang-core/tests/design_schema_drift.rs b/crates/ailang-core/tests/design_schema_drift.rs
index 0a6596b..4315df3 100644
--- a/crates/ailang-core/tests/design_schema_drift.rs
+++ b/crates/ailang-core/tests/design_schema_drift.rs
@@ -350,6 +350,82 @@ fn design_md_anchors_every_parammode_variant() {
     }
 }
 
+/// The anchor-presence check must scope its substring match to ```jsonc```
+/// (and ``` / ```json) fenced code blocks. The whole-file `.contains()` is
+/// fidelity-widened: a future edit can delete the canonical schema entry for
+/// a variant from inside a fenced block while leaving the anchor string
+/// mentioned in surrounding prose (e.g. a "design rationale" footnote, an
+/// inline reference like `the "t": "lit" form`, or a historical-note
+/// section). The drift test would silently pass, and downstream LLM authors
+/// would lose the canonical schema for a variant the AST still produces.
+///
+/// This test pins the property in two directions:
+///   1. A markdown string with the anchor present ONLY in prose (no fenced
+///      block) must report ABSENT under the scoped helper. The unscoped
+///      `.contains()` reports PRESENT for the same input — that mismatch is
+///      the bug surface.
+///   2. A markdown string with the anchor present ONLY inside a ```jsonc```
+///      block must report PRESENT under the scoped helper.
+///
+/// Gitea issue #10. The helper `anchor_in_jsonc_block` does not yet exist —
+/// this test is RED until `skills/implement` mini-mode adds it and re-routes
+/// the six existing call sites onto it.
+#[test]
+fn anchor_presence_check_is_scoped_to_jsonc_blocks() {
+    // Anchor mentioned only in prose — the false-pass surface. The whole
+    // sentence after the heading is plain markdown body text; no fenced
+    // block exists in this fixture at all.
+    let prose_only = r#"# Data model
+
+The `"t": "lit"` form used to be the canonical literal anchor. See
+historical note below.
+
+## Historical note
+
+Older drafts pinned the literal schema as `"t": "lit"`; the current
+schema uses a different shape (see commit log).
+"#;
+
+    // Anchor present only inside a fenced jsonc block — the true-positive
+    // surface. Prose outside the block does not mention the anchor.
+    let jsonc_only = r#"# Data model
+
+The literal form is canonical.
+
+```jsonc
+{ "t": "lit", "lit": Literal }
+```
+
+End.
+"#;
+
+    // Unscoped substring match — what the file does today. Both inputs
+    // report PRESENT, and that is the bug: prose-only must not count.
+    assert!(prose_only.contains(r#""t": "lit""#));
+    assert!(jsonc_only.contains(r#""t": "lit""#));
+
+    // Scoped helper — what the file must do. Prose-only is ABSENT,
+    // jsonc-only is PRESENT. This call is the RED: the helper does not
+    // yet exist, so this test fails to compile until `implement` mini-mode
+    // factors it out and re-routes the six call sites onto it.
+    assert!(
+        !anchor_in_jsonc_block(prose_only, r#""t": "lit""#),
+        "scoped helper must NOT count an anchor that lives only in prose"
+    );
+    assert!(
+        anchor_in_jsonc_block(jsonc_only, r#""t": "lit""#),
+        "scoped helper MUST count an anchor that lives inside a ```jsonc``` block"
+    );
+
+    // Live document: the helper must also report PRESENT for the live
+    // data-model.md (proves the helper does not over-narrow and break
+    // the existing six checks).
+    assert!(
+        anchor_in_jsonc_block(DATA_MODEL, r#""t": "lit""#),
+        "scoped helper must find live anchors inside data-model.md ```jsonc``` blocks"
+    );
+}
+
 /// Nested struct key anchors must be present in design/contracts/data-model.md.
 /// These keys appear inside `Suppress`, `ClassMethod`, `InstanceMethod`,
 /// and `Type::Forall` — they are not discriminators but they ARE part