harden against mib007 mention-path 500 (insert-then-error)

Smoke test found: any @handle in a comms message 500s AFTER the row is
inserted (agents.url_key column missing on this instance), so a retry
duplicates the message. Neutralise ALL '@' the bridge emits (incl. its
own @nir escalation, zwsp renders identically) and make comment HTTP
errors non-retryable best-effort drops; posts still retry.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YECpkAwQUwgu7NVy91R8fW

README.md

@@ -68,10 +68,15 @@ launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/ai.shre.pulse-bridge.pli
   only (the first `@handle` in a message that matches a workspace agent
   triggers an AI reply). There is no user mention/notification hook, so
   needs-you escalations are the literal text `@nir NEEDS YOU: …`.
-- **Mention trap defence:** all ledger-derived text has `@` neutralised
+- **Mention path is broken on this instance (found during smoke):** ANY
-  with a zero-width space so an `@ellie` in an item detail can never
+  `@handle` in a message makes `POST …/messages` **500 after the row is
-  trigger an AI reply in the feed. Only the bridge's own `@nir` (which
+  inserted** — the mention agent-lookup queries a nonexistent `url_key`
-  matches no agent, hence inert) is emitted raw.
+  column on `agents` (`comms.ts` ~line 226; error in mib007 stderr log).
+  Retrying such a failure duplicates the message. The bridge therefore
+  (a) zwsp-neutralises **every** `@` it emits, including its own
+  `@nir` escalation (renders identically in the UI), and (b) treats an
+  HTTP error on a *comment* as non-retryable (`DROP … comment not
+  retried` in the log) — comments are best-effort; posts still retry.
 - Updates for items that pre-date the bridge and were never seeded
   (closed before first run) are skipped with a log line — there is no
   post to comment on.

bridge.py

@@ -115,8 +115,11 @@ def neutralize(text: str) -> str:
     """Insert a zero-width space after '@' in ledger-derived text.
 
     mib007's comms route treats the FIRST '@handle' in a message as an agent
-    mention and fires an AI reply into the thread (e.g. @ellie). Ledger text
+    mention: a matching agent fires an AI reply into the thread, and on this
-    must never trigger that; only the bridge's own deliberate mention may.
+    instance ANY mention 500s after the insert (the agent lookup references a
+    nonexistent url_key column). So no bridge-emitted text may ever contain a
+    bare '@handle' — including the bridge's own needs-you escalation, which is
+    emitted zwsp-neutralised and renders identically in the UI.
     """
     return (text or "").replace("@", "@​")
 
@@ -151,8 +154,8 @@ def update_body(rec: dict) -> str:
     stage = rec.get("stage")
     note = neutralize(rec.get("note", ""))
     if kind == "needs-you":
-        lines.append(f"@{MENTION} NEEDS YOU: this item is now waiting on a human."
+        lines.append(neutralize(f"@{MENTION}") + " NEEDS YOU: this item is now"
-                     + (f" — {note}" if note else ""))
+                     " waiting on a human." + (f" — {note}" if note else ""))
         note = ""  # already included
     elif kind == "failed":
         where = f" at {stage}" if stage else ""
@@ -249,7 +252,14 @@ def comment(state: dict, ledger_id: str, content: str, what: str) -> None:
     if not post_id:
         log(f"SKIP {what} {ledger_id[:8]}: no mapped post (predates bridge)")
         return
+    try:
         msg = send_message(ensure_channel(state), content, thread_id=post_id)
+    except urllib.error.HTTPError as e:
+        # mib007 comms can 500 AFTER the insert (proven: mention-path crash),
+        # so retrying a failed comment risks duplicates. Comments are
+        # best-effort: log and move on instead of blocking the checkpoint.
+        log(f"DROP {what} {ledger_id[:8]}: HTTP {e.code} (comment not retried)")
+        return
     log(f"{what:6} {ledger_id[:8]} -> comment {msg['id']}")