Stop two tests reporting policy as breakage

Both predate this session's pipeline work — confirmed by running them at the
pre-session checkpoint — and both were asserting against premises the code had
deliberately moved past.

TestAnInterruptedStallIsStillAStall asserted that an identical failure three
times escalates, using "completeness" as its gate. completeness is in
escalationWouldNotHelp: it asks for text the run has already been given in full,
a dearer model cannot supply it any better, and record declines to escalate on
it whatever the tally. It now uses a gate with no exemption, and fails loudly if
that gate ever acquires one. This is the second test found today asserting a
general rule through a gate that is exempt from it.

TestAnUndocumentedMainIsAskedForRatherThanOnlyReported looked for a bare "main"
among the reported gaps. Gaps have been qualified with their file since
CL-20260803-141, because a generated workspace holds a stub main.go beside the
real one and the unqualified name describes both — so the test read the
qualification as the gate having stopped asking at all.

Change-Log: CL-20260803-166
Dev-Log: DL-20260803-176

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: monstercameron

CHANGELOG

@@ -50,9 +50,36 @@ Dev-Log:
 
 Entries
 -------
-Change-ID: CL-20260803-165
+Change-ID: CL-20260803-166
 Commit: pending
 Date: 2026-08-04
+Type: Fix -- two long-failing tests were asserting against premises the code had
+  deliberately moved past
+Request-or-TODO: Regression sweep after the rung 16 work
+Outcome: Both predate this session's pipeline work -- checked by running them at
+  the pre-session checkpoint -- and both were reporting a deliberate design
+  decision as a broken mechanism. TestAnInterruptedStallIsStillAStall asserted
+  that an identical failure three times escalates, using "completeness" as the
+  gate; completeness is in escalationWouldNotHelp, because it asks for text the
+  run has already been given in full and a dearer model cannot supply it any
+  better, so record declines to escalate on it whatever the tally. It now uses a
+  gate with no exemption and fails loudly if that gate ever acquires one -- the
+  same shape as the fix to TestARunRepeatingItselfIsEscalated earlier today, and
+  the second time this pattern has been found.
+  TestAnUndocumentedMainIsAskedForRatherThanOnlyReported looked for a bare
+  "main" in the reported gaps; gaps have been qualified with their file since
+  CL-20260803-141, because a generated workspace holds a stub main.go beside the
+  real one and the unqualified name describes both. It matches the qualified
+  name now.
+Affected-behavior: None. Both changes are to tests.
+Compatibility-or-migration: None.
+Verification: Both pass. Both were confirmed failing at 5577f21, before any of
+  this session's pipeline changes, so neither was caused by them.
+Dev-Log: DL-20260803-176
+
+Change-ID: CL-20260803-165
+Commit: bf60014
+Date: 2026-08-04
 Type: Fix -- the fuzz stage named the case it could not handle and then did the
   one thing guaranteed to fail
 Request-or-TODO: Ladder rung 16

DEVLOG

@@ -25,6 +25,27 @@ Next-safe-step:
 
 Entries
 -------
+Dev-Log: DL-20260803-176
+Date: 2026-08-04
+Status: Committed as CL-20260803-166
+Change-ID: CL-20260803-166
+Request-or-TODO: Regression sweep
+Goal: Clear two failures without mistaking a policy for a bug
+Assumptions: None. Both were run against a clean checkout at 5577f21 first, to
+  establish they were not caused by this session
+Decisions: Changed the tests, not the code, in both cases. The exemption for
+  completeness is documented with a reason and the qualification of gap names
+  was made deliberately to stop runs patching the wrong main.go; a test that
+  disagrees with either is the thing that is out of date
+Files-or-schemas: internal/coordinator/agent_convergence_test.go,
+  agent_refinable_test.go
+Validation: Both pass; the wider coordinator suite is unaffected
+Failures-or-discarded-approaches: None
+Known-limitations: This is the second stale test found asserting a general rule
+  through a gate that has a specific exemption. Both now guard against the gate
+  acquiring one, but nothing stops a third being written the same way
+Next-safe-step: Rung 17
+
 Dev-Log: DL-20260803-175
 Date: 2026-08-04
 Status: Committed as CL-20260803-165

internal/coordinator/agent_convergence_test.go

@@ -355,15 +355,27 @@ func TestDecompositionTellsTheRunToStopTryingItInOnePiece(t *testing.T) {
 // stuck on one gate indefinitely as long as something else interrupted now and
 // then. Worse, the interruptions came from the malformed-turn recovery added
 // to make runs survive model slips: one repair quietly disabled another.
+//
+// The gate is deliberately not "completeness", which is what rung 5 actually
+// failed. completeness is in escalationWouldNotHelp: it asks for text the run
+// has already been given in full, so a dearer model cannot satisfy it any
+// better and record declines to escalate on it whatever the tally. Written
+// against completeness this asserted the general rule through a gate that is
+// exempt from it, and reported a deliberate policy as a broken counter.
 func TestAnInterruptedStallIsStillAStall(t *testing.T) {
 	tracker := newConvergence(escalationSettings(3))
-	const stuck = "1 function has no doc comment: main"
+	const gate = "path-coverage"
+	const stuck = "1 changed line is executed by nothing: main.go"
+	if escalationWouldNotHelp[gate] || regressionProneGates[gate] {
+		t.Fatalf("%s now has an exemption, so this case no longer exercises "+
+			"the ordinary counting rule", gate)
+	}
 
-	tracker.record("completeness", stuck, stuck)
-	tracker.record("completeness", stuck, stuck)
-	// An unrelated failure. Nothing was learned about the doc comment.
-	tracker.record("assembly", "the loop refused a malformed turn", "")
-	decision := tracker.record("completeness", stuck, stuck)
+	tracker.record(gate, stuck, stuck)
+	tracker.record(gate, stuck, stuck)
+	// An unrelated failure. Nothing was learned about the uncovered line.
+	tracker.record("model-turn", "the loop refused a malformed turn", "")
+	decision := tracker.record(gate, stuck, stuck)
 	if decision.Escalated == "" {
 		t.Fatal("the third occurrence of an identical failure did not escalate " +
 			"because an unrelated failure happened in between")

internal/coordinator/agent_refinable_test.go

@@ -212,15 +212,22 @@ func TestAnUndocumentedMainIsAskedForRatherThanOnlyReported(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
+	// Matched on the qualified name, because that is what a gap is now called.
+	// A bare "main" is ambiguous exactly where it matters most: a generated
+	// workspace holds a stub main.go beside the real cmd/thing/main.go, so the
+	// unqualified name describes two declarations and the run has to guess. The
+	// assertion kept looking for the bare name and read the qualification as
+	// the gate having stopped asking at all.
 	asked := false
 	for _, name := range gaps.UndocumentedAtoms {
-		if name == "main" {
+		if name == "main" || strings.HasSuffix(name, ":main") {
 			asked = true
 		}
 	}
 	if !asked {
-		t.Error("the ledger reports an undocumented main and the gate never " +
-			"asks for one, so no attempt can close it")
+		t.Errorf("the ledger reports an undocumented main and the gate never "+
+			"asks for one, so no attempt can close it: %v",
+			gaps.UndocumentedAtoms)
 	}
 	if checkAtomDocumentation(worktree, newProducedFunctionCache(worktree)).Held {
 		t.Error("the ledger accepts an undocumented main the gate asks for, " +