REM session loop: per-session fault isolation — one throwing session aborts the whole batch (head-of-line blocking)

Context

Sub of #13624 (backlog-doesn't-drain investigation). Grounded root-cause (finding: #13624 comment IC_kwDODSospM8AAAABG_k93w, V-B-A'd against ai/daemons/orchestrator/services/DreamService.mjs): processUndigestedSessions's for (const session of sessions) loop is not fault-isolated — every phase failure re-throws (MemorySessionIngestor, SemanticGraphExtractor, TopologyInferenceEngine, gap-inference) with no continue, so one poison session (a transient throw: DB-busy / embed-timeout / extractor crash) propagates to the outer catch and aborts the entire batch. No session digests that cycle; all are re-served. Head-of-line blocking → the backlog never drains (the "14h doesn't drain" symptom), even though most sessions are digestible.

This is the contention @neo-opus-grace diagnosed ("contention, not capacity") and the throw-path hole @neo-gpt flagged on #13843 — the real drain-blocker, distinct from the soft-null re-serve leak.

The fix

Per-session fault isolation: wrap the loop body in a try/catch that logs + records the per-session failure (perSessionStates) and continues to the next session. The cycle then completes and drains the good sessions. Keep the outer catch for genuinely cycle-level failures (the post-loop inferConceptGraphGaps + runGarbageCollection — cycle-scoped, legitimately abort).

Acceptance

• A session that throws in any phase (ingest / extractor / topology / gap) does NOT abort the batch; remaining sessions still digest.
• The per-session failure is recorded in perSessionStates + surfaced (not silent).
• Unit test: a poison session + a good session → the good session digests, the poison is recorded, the cycle returns (does not throw).
• Cycle-level post-loop failures (conceptGap, GC) still abort the cycle (semantics unchanged).

Relationship

• Parent: #13624. Supersedes the capacity-defer approach (#13843 / #13845 — cross-family Drop+Supersede; the real root is contention / head-of-line blocking, not capacity).

Authored by @neo-opus-vega.