fix: restore persisted operator sessions after engine cache eviction #856

Problem

Custode persists compatible active operator conversation handles and seeds them when a provider process starts. Hot operator delivery passes only the arc and session mode. After 32 fresh scheduled/inbox arcs, the engines can evict the operator handle from their bounded cache, so the next operator message starts fresh despite Custode retaining a valid handle. A cold restart would restore that same handle.

The upstream prerequisite is released and verified on Hex and in tagged source:

• oban_claude 0.10.2, fix(agent): admit exact resume handle after cache eviction oban_claude#173
• oban_codex 0.7.1, fix(agent): admit exact resume handle after cache eviction oban_codex#50

Both engines accept resume_session_id: for a resume delivery without a fork. Admission occurs when the turn is delivered. A different live handle is refused; a supplied handle cannot redirect an answer away from its pending question's arc. LRU bounds remain unchanged.

Bounded implementation

1. Adopt the released dependency versions through the ordinary lockfile update.
2. Supply the exact persisted native handle for a compatible operator resume at the shared delivery boundary. Respect existing host/workspace/provider/configuration compatibility and rotation decisions. Do not send an old handle with fresh/fresh-fallback/fork requests.
3. Preserve queued-answer ownership from fix: preserve the question arc when replaying queued operator answers #853; a resume handle must not turn an answer into a new operator conversation.
4. Handle explicit admission conflicts without silently overwriting a live handle, losing the durable operator message, or replaying an effectful prompt into a fresh session.

Acceptance

• Deterministic controlled engine tests on both providers: establish an operator handle, finish 32 fresh sweep arcs, then send another operator message and assert the original native handle is resumed.
• Hot and cold paths agree on the selected compatible handle.
• Conflicting live handle, stale/rotated/configuration-incompatible handle, fresh delivery and pending-question cases retain explicit safe outcomes.
• Correlation IDs, durable queue behavior and restart reconciliation remain intact.
• No paid native runs needed for the initial regression; use controlled enqueue/result fixtures.

Keep missing-transcript classification separate: genagent/oban_claude#175 and genagent/oban_codex#52 remain open pending versioned CLI fixtures. This change does not establish safe automatic replay after an ambiguous provider failure.

Related: #853, #855. This is a current-runtime continuity fix, not a kernel or GenAgent adoption prerequisite.