submit_gateway_pending_mirror() matches an incoming approval back to its
queued producer entry by object identity (entry.data is approval) then by
approval_id. On the LOCAL backend neither holds: the installed hermes-agent
core delivers the approval to WebUI's notify callback as a COPY
(notify_cb(dict(entry.data))), so identity fails, and a local _ApprovalEntry
stamps a request_id but no approval_id, so the approval_id fallback misses.
The mirror is then created with no token, reconcile keeps the orphan, and
_session_has_pending_approval stays True after the entry is dropped — the
stale-approval-card dead-end (#4948 local variant): clicking a card whose
approval is gone returns a bare ok:false the UI renders as "Approval response
not accepted." with a stuck card.
Add a request_id fallback (gated on the local not-run_id path) that reunites
the notified copy with its queued entry via the request_id the core stamps
once on the source and preserves through the copy.
The four HTTP tests in test_approval_unblock.py were unfaithful to production:
they built _ApprovalEntry(approval) (stamping request_id onto a copy in
entry.data, not onto the source approval) then submitted the pre-stamp source.
They now submit dict(entry.data), matching what the core actually passes. The
assertions are unchanged; with the fidelity fix but the production change
reverted, the tests fail — so they detect the real regression, not vacuously.
This class is invisible on CI, which never installs the agent core, so both
approval test files skipif-skip entirely and their green never exercised these
paths.
Co-authored-by: n <a@n>
