zeph: Q96 -- absence cell + exhausted-stem recovery in q60_enforce.py

wizard-work/zeph/q60_enforce.py

@@ -29,9 +29,14 @@ What it does, on a checkout at a NAMED head:
      concatenation so no literal for it exists in this source; (b) FAIL
      diagnostics are masked (stem + length + site, never the full value),
      because a report quoting the fault re-commits it at the next head.
-     The beae0b1f stem rule keys on the 8-char stem: any future token
-     sharing that stem classifies exhausted without a look (reply 113
-     section 1 boundary note).
+     The beae0b1f stem rule keys on the 8-char stem, narrowed per reply
+     130 section 2 / interview 117 section 1.1 (Q96): a truncated token
+     sharing that stem skips the longer-string look and stays
+     UNRECOVERED-JUSTIFIED, but a full UUID-shaped value arriving under
+     that stem is reported RECOVERED with its site. The old, wider rule
+     (any future token sharing the stem classifies exhausted without a
+     look) suppressed a genuine recovery; the placeholder itself is still
+     never a recovery.
 
 Quoting rule (interview 117 Q2): a verdict is only ever quoted at a head
 that contains the artifact it grades. Green readings at heads predating
@@ -41,8 +46,19 @@ Exit status: 0 when no FAIL line is printed, 1 otherwise.
 ``--selftest`` builds fixture trees and asserts the negative cells: an
 unknown truncated token fails, a real complete-but-wrong value fails, the
 fixture value classifies TEST-FIXTURE, a self-quoting report file excluded
-from the population collapses the count, and the check over the checker's
-own directory passes (cell 7, the self-scan leg).
+from the population collapses the count, the check over the checker's
+own directory passes (cell 7, the self-scan leg), the checker's own
+source bytes do not contain the concatenated fixture value (cell 8, the
+absence cell, Q96), and a full value under the exhausted beae0b1f stem
+reports RECOVERED (cell 9, Q96).
+
+Correction, reply 130 probe D / interview 117 section 1.4: the
+2026-10-27 check-in credited cell 7 with proving that no UUID-shaped
+literal for the fixture exists in this source. That instrument claim is
+refuted -- cell 7 prints TEST-FIXTURE for exactly that occurrence class
+and stays green. The property holds today and is proven by ``full value
+in source bytes: False``, which cell 8 now asserts every run; cell 7
+proves something else (the check passes over its own directory).
 
 Usage:
   python3 q60_enforce.py --tree <checkout-path> [--exclude <relpath> ...]
@@ -184,6 +200,19 @@ def run_check(tree, excludes):
                         break
                 if longer:
                     break
+        else:
+            # Q96 narrowing: a truncated token under an exhausted stem
+            # skips the longer-string look, but a full value that arrives
+            # under that stem is a recovery, not an exhaustion. The
+            # placeholder (and the fixture, by symmetry) never counts.
+            for rel, text in contents.items():
+                for val in FULL_RE.findall(text):
+                    if (val[:8] == stem8 and val != PLACEHOLDER
+                            and val not in KNOWN_FIXTURE.values()):
+                        longer = (val, rel)
+                        break
+                if longer:
+                    break
         if longer:
             lines.append("TOKEN %s: RECOVERED-IN-TREE %s at %s" % (stem, longer[0], longer[1]))
         elif stem8 in KNOWN_EXTERNAL:
@@ -219,6 +248,19 @@ def selftest():
     """Fixture trees; each cell asserts a behaviour, not a tautology."""
     passed = []
 
+    # Cell 8 (Q96 absence cell): the concatenated fixture value must not
+    # appear in this file's own bytes. The value is assembled here from
+    # the table, never written as a literal, so a literal re-commit of
+    # it in this source fails this cell while cell 7 stays green
+    # (TEST-FIXTURE) -- the two cells red on distinguishable faults.
+    # Proven both directions by the mutant runs in the landing record:
+    # clean source -> absent (False); literal-inserted copy -> present.
+    fixture_value = "".join(KNOWN_FIXTURE.values())
+    with open(os.path.abspath(__file__), "rb") as fh:
+        assert fixture_value.encode("utf-8") not in fh.read(), \
+            "fixture value literal present in checker source"
+    passed.append("source-absence-passes")
+
     def check(tree, excludes, want_fail, label):
         _lines, failed = run_check(tree, excludes)
         assert failed == want_fail, "%s: want fail=%s" % (label, want_fail)
@@ -269,6 +311,18 @@ def selftest():
                    for l in lines), "fixture value did not classify TEST-FIXTURE at its site"
         check(d, ["report.md"], False, "fixture-value-passes")
         os.remove(os.path.join(d, "f.md"))
+        # Cell 9 (Q96 exhausted-stem flip): a full value arriving under
+        # the exhausted beae0b1f stem reports RECOVERED with its site.
+        # Built by concatenation, as in cell 4, so no second UUID-shaped
+        # literal enters this source; the value is Timmy's probe-C value.
+        with open(os.path.join(d, "g.md"), "w") as fh:
+            fh.write("beae0b1f-" "9f2c-4d1e-8a77-3c5b6e9d0a11\n")
+        lines, _ = run_check(d, ["report.md"])
+        assert any(l.startswith("TOKEN beae0b1f: RECOVERED-IN-TREE")
+                   and l.endswith("at g.md") for l in lines), \
+            "full value under exhausted stem did not report RECOVERED"
+        check(d, ["report.md"], False, "exhausted-stem-recovers")
+        os.remove(os.path.join(d, "g.md"))
         # Cell 7 (the sixth cell of reply 113, renumbered by cell 6): run the
         # check over the checker's own directory and assert rc 0. This is
         # the leg that was missing: five cells graded fixture trees, none