Reevaluate filesystem spillover for oversized tool results

Problem\n\nMecatl currently truncates model-facing tool output (or, for oversized structured MCP JSON, returns a fail-closed error). The cascade compactor can later replace old file-read bodies with a re-fetch pointer, but there is no general mechanism to preserve an oversized result outside the context window for selective follow-up.\n\nDisk spill was previously rejected because the no-fs profile and storage-free deployments must remain supported. That constraint does not necessarily rule out an optional capability: many sessions have a workspace filesystem.\n\n## Proposal to evaluate\n\nInvestigate an optional, generic oversized-tool-result spill facility:\n\n- When a session has an appropriate filesystem/workspace, persist an oversized result to a controlled session-scoped location and return a compact, model-visible reference plus retrieval affordance.\n- When the session is no-fs or its filesystem is unavailable, disable spill explicitly and retain the current truncation/fail-closed behavior.\n- Preserve the existing security and workspace invariants: no arbitrary path exposure, bounded storage/lifetime, safe cleanup, and no silent loss of structured-result semantics.\n- Establish how spilled data interacts with session persistence, compaction, event logs, delegated environments, and cloud-native/storage-free deployments.\n\n## Alternatives\n\nContinue relying only on remote pagination/filtering (including CallMcpWithQuery for MCP JSON) and truncation/compaction.\n\n## Relevant prior decision\n\nADR 0063 rejected unconditional scratch-file spill for portability reasons. This issue is to reconsider that decision in light of an optional filesystem capability, not to remove no-fs support.

Start by reading ADR 0063 and reviewing the current truncation, fail-closed, compaction, and CallMcpWithQuery behavior described in the issue. Done means a documented evaluation and decision covering optional filesystem use, no-fs behavior, security and workspace invariants, bounded cleanup, persistence, event logs, delegated environments, and storage-free deployments.