From 4ba99743733f0930291e57fa047014af9b48e6ed Mon Sep 17 00:00:00 2001
From: Aaron Sachs <898627+asachs01@users.noreply.github.com>
Date: Sun, 23 Aug 2026 02:14:22 +0000
Subject: [PATCH] fix(bus): widen check-stale-blockers precedent-citation
 exclude-list for stat-source citations

A real, correctly-cited PR can still be a false positive if it's cited as
the source of a supporting fact ("documented in-code by PR #NN as 13
verified against prod") rather than as the fix for the blocker. Distinct
shape from precedent-citation (that class isn't naming a blocker at all).

Bit the same task twice, two days apart (task_1786902033624, 2026-08-20
then 2026-08-22), both times resolved manually by the same peer -- the
tool has no memory of a prior resolution, so it re-litigates indefinitely.

Reuses the existing exclude-list mechanism (PRECEDENT_CITATION_CUE_REGEX,
2026-08-12) rather than new persisted state -- same failure shape, same
fix technique.
---
 src/bus/system.ts             | 14 ++++++-
 tests/unit/bus/system.test.ts | 70 +++++++++++++++++++++++++++++++++++
 2 files changed, 83 insertions(+), 1 deletion(-)

diff --git a/src/bus/system.ts b/src/bus/system.ts
index 3c24cba3a8..abbf05320d 100644
--- a/src/bus/system.ts
+++ b/src/bus/system.ts
@@ -249,11 +249,23 @@ const PR_REFERENCE_REGEX = /\bPR\s*#\s*(\d+)\b/gi;
 // narrow and nameable (precedent/example citation), so exclude it
 // specifically rather than narrowing the whole match surface.
 //
+// Widened 2026-08-22 (analyst, task_1786902033624 — a check-stale-blockers
+// sweep flagged the same false positive TWICE, two days apart, both times
+// resolved by the same peer): a genuinely real, correctly-cited PR can still
+// be a false positive if it's cited as the SOURCE OF A SUPPORTING FACT
+// ("documented in-code by conduit PR #1424 as 13 verified against prod")
+// rather than as the thing that resolves the blocker. This is a distinct
+// shape from precedent-citation (that class isn't naming a blocker at all;
+// this class names a real, relevant PR, just not as *the fix*) — added as a
+// second cue category on the same exclude-list mechanism rather than new
+// persisted state, since the failure mode (a citation-shape phrase near the
+// reference) and the fix (widen the window check) are identical in kind.
+//
 // No trailing \b: several cues end in a non-word char ("e.g."), and \b only
 // holds at a word/non-word transition — a trailing \b after "e.g." silently
 // never matches, since both the "." and the space after it are non-word.
 const PRECEDENT_CITATION_CUE_REGEX =
-  /\b(same (shape|pattern|approach|idiom) as|see .{0,10}for (the )?pattern|per the .{0,40}precedent|precedent|e\.g\.|for example|prior art)/i;
+  /\b(same (shape|pattern|approach|idiom) as|see .{0,10}for (the )?pattern|per the .{0,40}precedent|precedent|e\.g\.|for example|prior art|documented (in-code )?by|verified against|sourced from|per .{0,40}('s)? own (data|finding|number|stat))/i;
 
 // Some cues fully precede the reference ("same shape as ... PR #306"); one
 // straddles it ("see PR #12 for the pattern" — the reference sits INSIDE the
diff --git a/tests/unit/bus/system.test.ts b/tests/unit/bus/system.test.ts
index 8125db5b7b..161e341c5d 100644
--- a/tests/unit/bus/system.test.ts
+++ b/tests/unit/bus/system.test.ts
@@ -414,6 +414,28 @@ describe('Bus System', () => {
       expect(report.entries).toHaveLength(0);
     });
 
+    // task_1786902033624 (grower/analyst, 2026-08-20 then again 2026-08-22):
+    // a check-stale-blockers sweep flagged the SAME task twice, two days
+    // apart, both times a false positive — conduit PR #1424 is a real,
+    // correctly-cited reference, but it's cited as the source of a
+    // supporting statistic ("13 lapsed rows verified against prod"), not as
+    // the fix for this task's blocker. Distinct from precedent-citation
+    // (that class isn't naming a blocker at all; this one names a real,
+    // relevant PR, just not as the thing that resolves the block).
+    it('does not flag a PR mention cited as the source of a supporting fact, not the fix ("documented ... by PR #NN as", "verified against")', () => {
+      writeTask('myorg', {
+        id: 'task_stat_source',
+        title: 'Trialing rows with NULL current_period_end are invisible to the scheduler',
+        status: 'blocked',
+        description:
+          '14 of 15 trialing rows are lapsed, documented in-code by conduit PR #1424 as 13 verified against prod 2026-08-15.',
+      });
+
+      const report = checkStaleBlockers(testDir);
+
+      expect(report.entries).toHaveLength(0);
+    });
+
     it('still flags a genuine PR reference when "tool artifact" appears but not near that PR number', () => {
       writeTask('myorg', {
         id: 'task_unrelated_dismissal',
@@ -431,6 +453,54 @@ describe('Bus System', () => {
       expect(report.entries[0].detail).not.toContain('PR #55');
     });
 
+    // NOTE: the two mentions are kept far apart on purpose (separate
+    // sentences, no shared clause) — this exercise found that a stat-source
+    // cue sitting in the GAP between two close references (e.g. "...verified
+    // against prod. This one is blocked on PR#67...") can leak into the
+    // NEXT reference's before-window despite the existing prev-match-end
+    // clamp, because that clamp only protects against reaching before/into
+    // the previous match's own span, not the trailing prose after it. Not
+    // fixed here — pre-existing architecture limitation (the after-window
+    // has no symmetric next-match-start clamp), same class as the "e.g."
+    // trailing-period note above, just not previously triggered because no
+    // existing cue phrase naturally sits in that gap zone. Flagging in case
+    // a tighter mixed-citation case is ever reported for real.
+    it('still flags a genuine blocking PR mention even when a stat-source citation appears earlier, well-separated, in the same description', () => {
+      writeTask('myorg', {
+        id: 'task_stat_source_mixed',
+        title: 'ship the fix',
+        status: 'blocked',
+        description:
+          'Documented in-code by conduit PR #1424 as 13 verified against prod 2026-08-15. ' +
+          'Separately, this task itself cannot proceed until PR#67 merges.',
+      });
+
+      const report = checkStaleBlockers(testDir);
+
+      expect(report.entries).toHaveLength(1);
+      expect(report.entries[0].detail).toContain('PR #67');
+      expect(report.entries[0].detail).not.toContain('PR #1424');
+    });
+
+    it('does not flag other stat-source citation phrasings ("sourced from", "per PR #NN\'s own data")', () => {
+      writeTask('org-a', {
+        id: 'task_sourced_from',
+        title: 'reconcile the counts',
+        status: 'blocked',
+        description: 'The 41-org figure is sourced from PR #914.',
+      });
+      writeTask('org-a', {
+        id: 'task_own_data',
+        title: 'confirm the regression window',
+        status: 'blocked',
+        description: 'Per PR #200\'s own data, the window is 72 hours.',
+      });
+
+      const report = checkStaleBlockers(testDir);
+
+      expect(report.entries).toHaveLength(0);
+    });
+
     it('does not suppress a PR reference on the generic word "resolved" alone, only the narrow "tool artifact" cue', () => {
       writeTask('myorg', {
         id: 'task_generic_resolved',