fix(sessions): finalize abandoned session rows

[zaczzk]

Summary

• finalize old message-bearing session rows that have no end marker or activity heartbeat as terminal abandoned sessions
• run the bounded sweep at gateway startup and expiry cadence, including every served multiplex profile under its own config and database scope
• add state-layer and gateway regressions for eligibility, exclusions, idempotence, recovery behavior, configured idle windows, and cross-profile isolation

Validation

• uv run --frozen --extra dev pytest -q tests/hermes_state/test_orphan_gateway_session_repair.py tests/gateway/test_abandoned_session_sweep.py (20 passed)
• uv run --frozen --extra dev ruff check hermes_state.py gateway/run.py tests/hermes_state/test_orphan_gateway_session_repair.py tests/gateway/test_abandoned_session_sweep.py
• git diff --check

Scope

This PR completes the bounded state-layer repair. Broader exit-path lifecycle changes for ACP/bridge ownership remain a separate follow-up.

[andrexibiza]

Reviewed exact head ad0ea473b8846fdc059252b983c9426c043703b9 against base ec5e369fe68f4abb3188767c751d6166d2139f0a and current main 8e46e2c4c49391737e311aad12be829062364bac. The profile scoping and awaited AsyncSessionDB forwarding path check out. I found one P1 state-authority blocker.

P1 — missing heartbeat plus old start time is not authority to make a session permanently non-recoverable

GatewayRunner._sweep_abandoned_sessions() runs unconditionally at startup and on every expiry-watcher cadence, but passes only default_reset_policy.idle_minutes into the database sweep. It does not honor SessionResetPolicy.mode, per-platform reset overrides, current/sibling-process liveness, source ownership, or the newest transcript timestamp. SessionDB.finalize_abandoned_sessions() then selects on started_at < cutoff AND last_activity_at IS NULL and assigns terminal end_reason='abandoned', which this PR deliberately excludes from stale-route recovery.

The new real-DB multiplex regression proves the unsafe case rather than preventing it: it constructs GatewayConfig(multiplex_profiles=True) (default reset mode none), appends "hello" at the current time, backdates only started_at, clears last_activity_at, and expects both rows to be finalized. A persistent profile can therefore lose canonical routing to a transcript with a fresh message merely because heartbeat metadata is absent. This is a permanent session-boundary transition, not harmless cleanup metadata.

That conflicts with the merged startup-sweep contract in #93430: its #65422 / @halaprix lineage requires both started_at and newest message time to be stale, limits the source set, excludes live in-memory IDs, and uses recoverable startup_orphan_reap; #93430 also preserves #65478 / @hansai-art for the WS-sidecar/Desktop coverage. Open #92744 by @daglimmer is the complementary stale-open backlog surface and already uses the shared freshest-of-heartbeat/latest-message/start recency rule while archiving reversibly. #56603 by @adityachaudhary99 is the adjacent actual-expiry finalizer and explicitly leaves mode: none sessions alone until a real expiry event.

Required before merge:

• Bind this mutation to explicit policy/ownership authority: honor mode: none and per-source overrides, or introduce a separate opt-in maintenance policy instead of borrowing idle_minutes while ignoring its mode.
• Select and revalidate on canonical real recency (the existing _sql_session_last_active shape), not started_at alone.
• Do not assign a permanently non-recoverable reason without positive owner-death evidence; ambiguous legacy backlog should use the recoverable/archive lane.
• Add adversarial regressions proving that default mode: none, old-start + fresh-message + null-heartbeat, per-platform overrides, and a live owner all survive, while a genuinely stale/dead row finalizes.

Exact-head hosted CI/status evidence is absent at review time; the PR reports 20 focused local passes plus Ruff. Even a hosted green run would not close this blocker while the real-DB regression continues to assert the unsafe outcome.