Stop a presentation bound destroying the attempt it describes

The test-proposal reader bounds each reproduction step at 300 characters and
throws past a 25% tolerance window. Nothing reads that value: the call discards
its result, and both downstream consumers already trim the same field to the
same 300. So the throw cannot protect a caller; it only destroys the attempt.

Measured over 838 complete logs: 4 attempts died on this bound at 380-398
characters, every one descriptive prose, and build 15105015 spent its final
attempt on one. The window's own comment justifies 25% by citing overages of 15
and 13 characters (~1.02x); live overages sit at 1.27-1.36x, systematically just
outside, so the justification is stale.

The publisher's sibling validator had already reached this conclusion for the
same field. The authoring reader never learned it.

Chosen over widening the window because the window is shared: for a -match
guard a silent trim can remove a match and disable a check without saying so.
The strict siblings are named in the test and stay strict - the verifier matches
the failure signature against printed assertion text, and three guards match the
trigger and behavior fields.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 735ac9a2-7bec-4baa-ad19-c298e5bc795a

Author: Copilot CI

.github/scripts/Replicate-Issue.Tests.ps1

@@ -3711,6 +3711,59 @@ PS-STEP-FAILED: step 3 did not find its target
         $script:Source | Should -Match 'Test reproduction step \$\(\$stepIndex \+ 1\)'
     }
 
+    It 'reserves the prose path for the proposal field that nothing reads' {
+        # The identity of the opted-in site is the invariant, not the count: the fix
+        # panel legitimately marks eight of its own model-written fields prose, and a
+        # ninth may be added. What must never drift is the strict set below. Each of
+        # these is read: the verifier -match's the failure signature against the text
+        # the assertion prints; the orientation, safe-area and visual guards -match the
+        # trigger and behavior fields; and the sandbox step feeds the timing-sensitive
+        # -notmatch check. A silent trim there can remove a match and disable a guard
+        # without saying so, or add one and throw spuriously.
+        $ast = [System.Management.Automation.Language.Parser]::ParseFile(
+            $script:ScriptPath, [ref]$null, [ref]$null)
+        $calls = $ast.FindAll({
+                param($node)
+                $node -is [System.Management.Automation.Language.CommandAst] -and
+                $node.GetCommandName() -eq 'ConvertTo-BoundedAgentLine'
+            }, $true)
+
+        $prose = @{}
+        foreach ($call in $calls) {
+            $elements = @($call.CommandElements)
+            $description = $null
+            for ($i = 0; $i -lt $elements.Count - 1; $i++) {
+                if ($elements[$i] -is [System.Management.Automation.Language.CommandParameterAst] -and
+                    $elements[$i].ParameterName -eq 'Description') {
+                    $description = $elements[$i + 1].Extent.Text.Trim("'", '"')
+                }
+            }
+            if ($null -ne $description) {
+                $prose[$description] = @($elements | Where-Object {
+                        $_ -is [System.Management.Automation.Language.CommandParameterAst] -and
+                        $_.ParameterName -eq 'Prose'
+                    }).Count -gt 0
+            }
+        }
+
+        $stepKey = @($prose.Keys | Where-Object { $_ -like 'Test reproduction step*' })
+        $stepKey.Count | Should -Be 1 -Because 'the test reader bounds its steps in one place'
+        $prose[$stepKey[0]] | Should -BeTrue -Because 'nothing reads a test reproduction step'
+
+        $guarded = @(
+            'Sandbox reproduction step'
+            'Test expected behavior'
+            'Test observed behavior'
+            'Reported issue trigger'
+            'Automated test trigger'
+            'Test expected failure signature'
+        )
+        foreach ($field in $guarded) {
+            $prose.ContainsKey($field) | Should -BeTrue -Because "$field must still be bounded"
+            $prose[$field] | Should -BeFalse -Because "a guard matches against $field"
+        }
+    }
+
     It 'neutralises a logging command it echoes back to the agent' {
         $message = { ConvertTo-BoundedAgentLine `
                 -Value 'a ##vso[task.setvariable variable=x]y' `

.github/scripts/Replicate-Issue.ps1

@@ -5693,10 +5693,22 @@ function Read-TestProposal {
         throw 'The test proposal must contain 1-10 reproduction steps.'
     }
     for ($stepIndex = 0; $stepIndex -lt $steps.Count; $stepIndex++) {
+        # -Prose, unlike every sibling in this reader. A test reproduction step is
+        # read by nothing: Publish-ReplicationPR renders it into the PR body and no
+        # guard matches against it, unlike the sandbox reader's steps (which feed the
+        # timing-sensitive check) or the trigger and behavior fields below (which feed
+        # the orientation, safe-area and visual guards). Both stages downstream already
+        # trim it to exactly 300 without complaint - the manifest writer via
+        # ConvertTo-ReplicationSafeLog, and the publisher via ConvertTo-NormalizedPlainStep
+        # -Prose - and this call discards its result, so refusing here cannot protect
+        # anything a caller reads. It only destroys the attempt. Measured over 838
+        # complete logs: 4 attempts died on this bound at 380-398 characters, every one
+        # of them descriptive prose, and build 15105015 spent its final attempt on one.
         $null = ConvertTo-BoundedAgentLine `
             -Value $steps[$stepIndex] `
             -Description "Test reproduction step $($stepIndex + 1)" `
-            -MaximumLength 300
+            -MaximumLength 300 `
+            -Prose
     }
     $null = ConvertTo-BoundedAgentLine -Value $proposal.expectedBehavior -Description 'Test expected behavior'
     $null = ConvertTo-BoundedAgentLine -Value $proposal.observedBehavior -Description 'Test observed behavior'