fix(test): context-restore assertion is evidence-based, not prose-matching

The test failed twice per run in TWO CI cycles while passing locally
4/4: the prompt said 'present the content' and the check grepped the
FINAL message for exact phrases — local runs quoted the file, CI runs
paraphrased ('the most recent context is from branch-b...') and the
substring check lost the coin flip.

- prompt now demands machine-checkable output: the newest file's
  '## Working on:' heading VERBATIM + a literal 'RESTORED: <filename>'
  marker (the mtime-scramble and cross-branch subject matter untouched)
- assertion ordered strongest-first: RESTORED marker → legacy content
  phrases → tool-call corroboration (Read/Bash input naming the newer
  file, credited ONLY when the older file was never read — a
  both-files run must still present the right one)
- the older-file negative got STRONGER: an explicit RESTORED marker
  naming the older file fails even if wintermute words appear elsewhere
- sibling scan: context-recovery-artifacts got the additive prompt-side
  treatment only (quote the matched literals verbatim); its lenient
  1-of-6 assertion deliberately unchanged

3/3 consecutive local green with all evidence classes firing
(marker=true, content=true, toolNewer=true, toolOlder=false).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test/skill-e2e-session-intelligence.test.ts

@@ -163,7 +163,8 @@ IMPORTANT:
   Replace any references to ~/.claude/skills/gstack/bin/ with ./bin/ when running commands.
 - Do NOT use AskUserQuestion.
 - Just run the preamble bash block and report what you see.
-- Look for "RECENT ARTIFACTS" and "LAST_SESSION" in the output.`,
+- Look for "RECENT ARTIFACTS" and "LAST_SESSION" in the output.
+- In your final message, quote VERBATIM (copy exactly, do not paraphrase) any output lines containing "RECENT ARTIFACTS" or "LAST_SESSION".`,
       workingDirectory: workDir,
       maxTurns: 10,
       allowedTools: ['Bash', 'Read', 'Write', 'Edit', 'Grep', 'Glob'],
@@ -337,7 +338,10 @@ IMPORTANT:
 - Look in ${checkpointDir} for saved context files.
 - Current branch is "main" — do NOT filter by current branch. Load across all branches.
 - The newest file by YYYYMMDD-HHMMSS prefix is the canonical "most recent". Filesystem mtime has been scrambled — do not use it.
-- Do NOT use AskUserQuestion. Just present the content of the newest file.`,
+- Do NOT use AskUserQuestion. Just present the content of the newest file.
+- Your final message MUST end with these two lines (they are machine-checked — copy them exactly, do not paraphrase):
+  1. The newest file's "## Working on:" heading line, VERBATIM as it appears in that file.
+  2. A literal marker line: RESTORED: <filename of the newest file>`,
       workingDirectory: workDir,
       maxTurns: 8,
       allowedTools: ['Bash', 'Read', 'Grep', 'Glob'],
@@ -349,8 +353,34 @@ IMPORTANT:
     logCost('context-restore', result);
 
     const output = result.output ?? '';
-    const loadedNewer = output.includes('newer wintermute work') || output.includes('wintermute integration');
-    const loadedOlder = output.includes('old work') && !output.includes('newer');
+
+    // Evidence-based checks. CI receipts showed the agent finding + reading the
+    // RIGHT file, then paraphrasing the final message ("the most recent context
+    // is from branch-b...") — exact-substring checks over stochastic prose
+    // flaked. Three signal classes, strongest first:
+    //   1. Machine-checkable output contract (prompt demands the verbatim
+    //      "## Working on:" heading + a "RESTORED: <filename>" marker line).
+    //   2. Lenient content echo (legacy): distinctive newer-file phrases.
+    //   3. Tool-call corroboration: a tool call whose INPUT names the newer
+    //      file. Corroboration only — if the agent also read the OLDER file,
+    //      tool evidence is void and the final output must present the newer.
+    const newerFileName = '20260202-130000-newer-wintermute-work';
+    const olderFileName = '20260101-120000-old-work';
+    const newerMarker = new RegExp(`RESTORED:.*${newerFileName}`, 'i').test(output);
+    const olderMarker = new RegExp(`RESTORED:.*${olderFileName}`, 'i').test(output);
+    const newerContent = output.includes('newer wintermute work') || output.includes('wintermute integration');
+    const outputPresentsNewer = newerMarker || newerContent;
+
+    const toolInputs = result.toolCalls.map(tc => JSON.stringify(tc.input ?? {}));
+    const toolReadNewer = toolInputs.some(input => input.includes(newerFileName));
+    const toolReadOlder = toolInputs.some(input => input.includes(olderFileName));
+
+    // Presenting the OLDER file fails: an explicit RESTORED marker naming it,
+    // or older-file content with no newer-file presentation alongside.
+    const loadedOlder = olderMarker || (output.includes('old work') && !outputPresentsNewer);
+    // Tool evidence counts only when the older file was never read: a run that
+    // reads BOTH files must present the NEWER one in the final output to pass.
+    const loadedNewer = outputPresentsNewer || (toolReadNewer && !toolReadOlder);
     const exitOk = ['success', 'error_max_turns'].includes(result.exitReason);
 
     recordE2E(evalCollector, 'context-restore loads latest', 'Session Intelligence E2E', result, {
@@ -361,6 +391,6 @@ IMPORTANT:
     expect(loadedNewer).toBe(true);
     expect(loadedOlder).toBe(false);
 
-    console.log(`context-restore: loadedNewer=${loadedNewer}, loadedOlder=${loadedOlder}`);
+    console.log(`context-restore: loadedNewer=${loadedNewer} (marker=${newerMarker}, content=${newerContent}, toolNewer=${toolReadNewer}, toolOlder=${toolReadOlder}), loadedOlder=${loadedOlder}`);
   }, CAPTURE_MS);
 });