fix(issue-23): the remote audit record names the dispatch it vouches for; a record without identity fails closed

Review round 4 (planner-reviewer-claude-01 B-1, planner-reviewer-codex-01
P2, reached independently): dispatch_verified is a boolean. It says the
subscriber's token check ran when the record was written, but names no
dispatch, the token being masked by the audit redactor and attempt_id not
written. events.ndjson is appended to across a job's attempts, so a started
verified under attempt 1 satisfied every later attempt's wait - the very
misattribution R1 forbids, alive on the remote path while the local path
(data.dispatch_token in events.log) was already immune. codex reproduced it
with the real signed producer, the real _Watcher, the real audit writer and
the real operator re-dispatch; both waits answered 0 where 1 was due, even
when attempt 2's subscriber had correctly dropped attempt 1's replayed packet.

The subscriber now writes dispatch_fingerprint = SHA-256 of the token it
verified, and attempt_id for readers. With a token demanded, the predicate
accepts an audit record only if it is dispatch_verified AND its fingerprint
equals SHA-256 of that token; a record with no fingerprint - written before
the field existed - names nothing and is rejected on that path. Without a
token, any received started counts, as before.

Fingerprint over attempt_id as the identity, weighing the two proposals: the
local path already compares the token, so hashing it keeps one identity axis
for both sources; the token is a uuid4 nonce (registry.py:141), 128 bits, so
its SHA-256 is not recoverable and it is a CAS fence rather than the HMAC
secret in the first place; and attempt_id cannot tell apart a token
re-minted within one attempt, where the token can. attempt_id is recorded
because it is cheap and readable, not relied on.

W7 pins codex's five cases end to end - current dispatch accepted, an old
attempt's record rejected after a real re-dispatch, still rejected when the
stale packet is replayed and dropped, the new dispatch's own started
accepted, a bad HMAC never recorded - with create's wait and the wrapper's
wait required to agree on each. W8 pins the fail-closed legacy record and
the untokened compatibility path. W5 gains the no-fingerprint and
other-dispatch cases; W6 checks the real _Watcher writes the fingerprint and
never the token. S7e's status reset had no reason left and is gone. codex's
own reproduction file, unmodified, passes 5/5 against this tree.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

.agents/skills/lib.sh

@@ -2233,8 +2233,15 @@ except Exception:
 #    `started` it received, the only evidence a REMOTE agent (no shared disk)
 #    can leave. The subscriber verifies the HMAC and drops any event whose
 #    data.dispatch_token differs from the job's expected token before it
-#    writes this record, and marks it dispatch_verified when that check ran;
-#    with <dispatch_token> given, only a dispatch_verified record counts.
+#    writes this record (dispatch_verified), and writes dispatch_fingerprint
+#    = SHA-256 of that token. The file is appended to across a job's
+#    attempts, so the flag alone would let an earlier attempt's record
+#    acknowledge a later dispatch (review round 4): with <dispatch_token>
+#    given, a record counts only if it is dispatch_verified AND its
+#    fingerprint equals SHA-256(<dispatch_token>). A record with no
+#    fingerprint (written before this field existed) names no dispatch and
+#    is rejected on that path - fail closed. Without a token, any received
+#    started counts, as before.
 #
 # Never the registry status: a record leaves `pending` when another writer
 # closes it or another dispatch claims it. Lines are parsed as JSON, never
@@ -2247,8 +2254,9 @@ delegate_started_recorded() {
   local audit="$logs_dir/$job_id/events.ndjson"
   [ -f "$ev_log" ] || [ -f "$audit" ] || return 1
   EV_LOG="$ev_log" EV_AUDIT="$audit" EV_TOK="$tok" python3 -c '
-import json, os, sys
+import hashlib, json, os, sys
 tok = os.environ.get("EV_TOK", "")
+fp = hashlib.sha256(tok.encode("utf-8")).hexdigest() if tok else ""
 
 def lines(path):
     if not os.path.isfile(path):
@@ -2271,7 +2279,9 @@ for e in lines(os.environ["EV_LOG"]):
 for e in lines(os.environ["EV_AUDIT"]):
     if e.get("event") != "received" or e.get("source_event") != "started":
         continue
-    if not tok or e.get("dispatch_verified") is True:
+    if not tok:
+        sys.exit(0)
+    if e.get("dispatch_verified") is True and e.get("dispatch_fingerprint") == fp:
         sys.exit(0)
 sys.exit(1)
 ' 2>/dev/null

.agents/skills/lib_py/delegation/job_subscriber.py

@@ -24,6 +24,7 @@ Exit codes:
 from __future__ import annotations
 
 import argparse
+import hashlib
 import json
 import logging
 import queue
@@ -279,13 +280,19 @@ class _Watcher:
         # that survives defensive parsing is recorded here, including ones a
         # different host published. This is the external-observer record that
         # backstops the publisher's own "published" line if it never wrote one.
-        # `dispatch_verified` says this record is attributable to the job's
-        # current dispatch: the token check above ran (the registry record had
-        # a dispatch_token and the event's data.dispatch_token matched it) and
-        # HMAC verification passed before that. The token itself is never
-        # written - the audit redactor masks token-named keys - so this flag is
-        # what a local reader (delegate_started_recorded) can trust for a
-        # remote agent's `started`, which has no line in events.log.
+        # Which dispatch this record vouches for. `dispatch_verified` says the
+        # token check above ran (the record had a dispatch_token and the
+        # event's data.dispatch_token matched it) after HMAC verification; on
+        # its own it names no dispatch, and this file is appended to across a
+        # job's attempts, so a reader waiting for a later dispatch would take
+        # an earlier attempt's record as its acknowledgement (review round 4).
+        # `dispatch_fingerprint` is SHA-256 of the token that was verified -
+        # the same identity the local path compares in events.log, minus the
+        # token itself, which the audit redactor masks by key name and which a
+        # 128-bit uuid4 nonce is not recoverable from. `attempt_id` is kept for
+        # readers; it is not the identity, since a token can be re-minted
+        # within an attempt. A reader that demands a token must match the
+        # fingerprint; a record without one proves nothing about identity.
         mqtt_common.append_event(jid, {
             "event": "received",
             "source_event": payload.get("event"),
@@ -294,6 +301,10 @@ class _Watcher:
             "timestamp": payload.get("timestamp"),
             "detail": payload.get("detail", ""),
             "dispatch_verified": expected_disp_tok is not None,
+            "dispatch_fingerprint": (hashlib.sha256(expected_disp_tok.encode("utf-8")).hexdigest()
+                                     if expected_disp_tok else None),
+            "attempt_id": (data.get("attempt_id") if isinstance(data, dict) and data.get("attempt_id") is not None
+                           else self.attempts.get(jid)),
         })
         self.events.put((msg.topic, payload))