## Context

Earlier changes introduced channelized log ingest and artifact transfer contracts, but the Platform service keeps log batches, artifact transfer state, and artifact payload bytes in process memory. Metrics are currently generated from a live snapshot, backups are not represented, and Run remote access execution is a success-only placeholder. The repository already has file and MySQL metadata snapshots, signed Run requests, endpoint/session fencing, scoped plugin permissions, and safe frontend projections. This change extends those boundaries without replacing them.

## Goals

- Make restart behavior explicit and testable for logs, artifacts, metrics, and backups.
- Keep large bodies outside lightweight job/control payloads and expose only bounded, redacted projections to users, plugins, and platform_web.
- Make remote adapters declaration-driven and auditable, with cancellation and lease/attempt fencing handled by the existing job channel.
- Preserve independent channel priorities so slow artifact or adapter work cannot delay control heartbeat, job ack/result, or log acknowledgement.

## Decisions

### Decision 1: File-backed durable bodies behind service interfaces

Log segments, artifact chunks, and completed artifact content use owner-only files under configured private roots. Metadata and cursors remain in the existing repository snapshot (file or MySQL). Atomic temp-file rename, bounded reads, checksum verification, and startup reconstruction keep writes recoverable. Memory stores remain available for unit tests.

### Decision 2: Transfer manifests are the recovery source of truth

Artifact transfer sessions persist an idempotency key, owner scope, direction, size/chunk limits, received indexes, and final checksum. A restart reloads incomplete manifests and reconstructs next-missing state without exposing payload paths. A chunk is removed from the Run queue only after an acknowledgement for the exact transfer/artifact/index.

### Decision 3: Retention is bounded and deterministic

Log streams carry bounded retention count/age metadata; metric samples and backup records have configurable maximum records/bytes and oldest-first pruning. Pruning emits an audit record and never changes an acknowledged log sequence or an available artifact checksum. Recovery marks interrupted backups/receipts as recoverable failure rather than claiming success.

### Decision 4: Metrics and backups are append-only records with safe projections

Metric samples store server/run identity, timestamp, bounded numeric values, and source. Backup records store logical scope, artifact reference, checksum, size, state, and recovery/audit status. Host paths, credentials, sockets, PID, session/lease tokens, secret refs, and hashes used for fencing are excluded from response DTOs.

### Decision 5: Remote adapters are a constrained registry, not a command tunnel

The Platform registers adapter declarations from an installed plugin/runtime profile and authorizes a request only when owner/admin scope, server instance, selected endpoint, declared capability, and target allowlist all match. Run accepts a typed adapter kind and logical target key, validates timeout and retry bounds, checks context cancellation before and during work, and returns a safe result reference. Shell source, arbitrary command vectors, raw socket addresses, unapproved hosts, and embedded credentials are rejected. Existing Job attempt/lease/session fencing remains authoritative.

### Decision 6: Lightweight routes stay isolated

Control and job routes continue to reject log/artifact payload fields. Log ingest and artifact transfer use separate clients/queues. Remote adapter work is scheduled as a job and its result is metadata-only; it cannot write through control or log endpoints. Tests exercise interleaving and blocked/slow operations with bounded deadlines.

## Data Flow

1. Run appends a bounded log batch or artifact chunk to its owner-only spool/queue before upload.
2. A low-priority uploader sends the batch/chunk on its dedicated route; the Platform validates signature, endpoint/session, owner scope, sequence/range/checksum, and idempotency before durable append.
3. Platform persists metadata and body/manifests atomically, returns an acknowledgement, and the Run removes only the acknowledged item.
4. Metrics and backup records are written through bounded service methods, pruned deterministically, and queried through authorized safe DTOs.
5. A declared remote adapter request becomes a fenced job. Run executes only the typed adapter implementation, returns a bounded status/result ref, and Platform audits/project results after terminal fencing.

## Non-Goals And Follow-Ups

- No dependency installation, Run self-update, client-manager lifecycle, plugin lifecycle, production scaling/alerts, external object stores, arbitrary FTP/rsync/DB/RCON network access, or real AI provider integration.
- Native OS process birth tokens and distributed transaction guarantees remain outside this change.

## Rollback

The additive protocol and repository fields are backward compatible. Removing the new capabilities stops advertising durable/adapter features and leaves old metadata untouched; incomplete transfer manifests remain private and can be retried by a compatible release.