fix(email): preserve thread recipients and cursor state

[aviyashchin]

What

• Persist the IMAP UID cursor only after an inbound message is safely admitted.
• Preserve same-thread In-Reply-To/References and reply-all recipient sets while removing the bot itself.
• Support explicit Email subject and CC through the existing send-message tool.
• Keep configured always-CC recipients without duplicating To/Cc addresses.

Why

Advancing the cursor before processing can lose a message across restart, and reducing a reply to only the sender drops legitimate thread participants. The send path also could not express an explicit subject/CC even though the adapter supports email delivery.

Validation

• scripts/run_tests.sh tests/gateway/test_email.py tests/gateway/test_email_robustness.py tests/gateway/test_email_charset_fallback.py tests/gateway/test_email_secret_scope.py tests/tools/test_send_message_tool.py tests/tools/test_send_message_target_parse.py
• 125 passed, 0 failed on macOS arm64
• Production canary: authenticated Gmail inbound to the live adapter produced a two-message same thread; the reply carried In-Reply-To, preserved the original To and Cc recipients, and returned the exact requested token.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Reviewed the diff. Solid upgrade on both axes. The durable UID cursor fixes a real restart-safety hole (another client marking mail SEEN while Hermes is down silently swallowed those messages under the UNSEEN search), with write-after-dispatch ordering proven by the failure test, atomic tmp+replace persistence, and an `int(uid) <= cursor` guard covering the `X:*` inclusive-range quirk. Recipient preservation is similarly complete: getaddresses-based extraction with dedup, thread-context carrying To/Cc into replies (self dropped, extras de-duped into Cc), always-cc/reply-all-senders config plus per-send cc/subject plumbed through both the adapter path and the standalone sender, with tests pinning each combination.

Four things worth attention:

• plugins/platforms/email/adapter.py:~600 (_load_uid_cursor) — the stored `imap_host` is written but never validated on load. UIDs are scoped to a specific server's mailbox; if the same address is later pointed at a different host (migration, provider switch), the stale high-water mark silently suppresses everything below it on the new host. Suggestion: compare `payload["imap_host"]` against `self._imap_host.lower()` and discard the cursor on mismatch (log it).

• One-time upgrade gap: on the first run after this change `_last_processed_uid` is None and connect bootstraps it to the current newest UID — so mail that arrived while the previous version was down (which UNSEEN would have caught) is skipped exactly once. Worth a line in the changelog; alternatively seed the cursor from the lowest unseen UID on that first bootstrap.

• Head-of-line blocking is now durable. Since the cursor only advances post-dispatch, one persistently failing message (poison attachment, handler crash) both stops cursor progress and, if `_dispatch_message`'s exception escapes the poll loop unchecked, halts polling altogether. Confirm the poll task wraps `_check_inbox` failures; consider a retry-cap that logs-and-skips a repeatedly failing UID past N attempts so one bad message can't freeze the inbox.

• Two sources of truth for external correspondents. The adapter accepts `extra.external_correspondents` or `EMAIL_EXTERNAL_CORRESPONDENTS", while the gateway-side isolation (the companion mechanism for skipping memory/persona) reads only the config key. A user configuring via the env var gets the "[External correspondent email]" tag but full operator context. Unify on one lookup path.

No blocking issues found — the cursor work in particular is the right fix.