feat: BYO clients ship their own container logs; drop the CLI's detached shipper

[7174Andy]

Summary

• BYO clients now ship their own container logs to the allocator from inside the container: start.sh routes fd 5 (every service's tagged output) through a new ship_logs worker in the client package when SHIP_LOGS=1, which lablink client register writes for every BYO shape.
• Closes the gap where hand-off clients (register --no-run-locally, e.g. Run:AI workloads) had no log shipper at all — the container is its own PID 1 with no docker daemon to tail, so their allocator logs pages stayed permanently empty.
• Deletes the CLI's detached log_shipper.py (441 lines + 673 test lines), its PID/state-file machinery, Docker.follow_logs, and the psutil dependency — net −1,284 lines.

Changes Made

Client package

• New lablink_client_service/ship_logs.py + ship_logs console script: reads fd 5 from stdin, writes each line to container stdout first (so docker logs output is byte-identical), queues a UTC-stamped copy in a bounded drop-oldest deque, and ships batches (50 lines / 15 s) to POST /api/vm-logs/<VM_NAME> from a separate thread. SIGTERM/EOF do a single-attempt final flush so a graceful docker stop ships the tail; missing env degrades to pure passthrough.
• start.sh: when SHIP_LOGS=1, fd 5 feeds a supervisor loop that respawns a crashed worker (the kernel pipe buffer absorbs the gap) and exec cats after 3 failures — worst case is exactly today's plumbing, never a frozen container.

CLI package

• register writes SHIP_LOGS=1 into the env file unconditionally (run-locally via --env-file, hand-off via the paste block); the shipper spawn/revive/stop machinery is deleted and _resume shrinks to container-only. CONTAINER_NAME moves to register.py.
• client doctor's shipper check becomes docker exec lablink-client pgrep -f ship_logs (pgrep verified present in the published image).
• docs/reference/cli.md's "in-container shipper" wording is now literally accurate; docs/cli/byo-clients.md documents the hand-off paste requirement.

Design Decisions

• Read the source stream, no file intermediary. An earlier tee-to-file design was rejected because it rebuilt the architecture of the fix: resolve VM log shipping data loss from Docker json-file driver stall #304 json-file tail stall (writer freezes silently, shipper tails a dead file forever) and made tee a single point of failure for all container logging.
• Passthrough-first, fail-open. The worker sits in the logging path, so the read loop never blocks on the network (bounded queue, shipping thread) and every failure mode degrades to plain passthrough — logs visible in docker logs / runai workspace logs, absent from the allocator, loudly visible either way.
• One shipper for all BYO shapes instead of a hand-off-only patch: the CLI's detached shipper was the flakiest component (a background process on operator laptops that died silently in live testing) and docker logs is the same stream the container sees internally.
• 4xx drops the batch instead of killing the worker (unlike the old CLI shipper): nothing respawns this worker for shipping-only failures, so a transient 401 during allocator DB warmup must not permanently kill the only log channel.
• Log group is container-docker — the allocator's existing -docker suffix routing lands it in the docker_logs column with zero allocator changes.

Release ordering

The client image must be published with the ship_logs entry point before a new CLI release reaches operators — a new CLI plus an old image means BYO clients log locally but nothing reaches the allocator (old images simply ignore SHIP_LOGS). AWS clients are untouched (user_data's log_shipper.sh remains the host-side shipper there, including cloud-init logs).

Testing

• 11 new client tests (test_ship_logs.py): passthrough byte-fidelity and ordering, timestamping, drop-oldest overflow, batch drain, retry/drop semantics, single-retry final flush, interval + stop flushes.
• CLI tests reworked: SHIP_LOGS=1 asserted in env file and hand-off printout, doctor's new exec-based check, resume path without shipper revival; stale shipper tests deleted.
• Full suites: client 178 passed; CLI 777 passed (12 pre-existing local-only ANSI rendering failures, unrelated).
• Live shell simulation of the supervisor: stub worker crashing every 2 lines → all 12 lines passed through exactly once across 3 respawns, then the cat fallback engaged.
• Real ship_logs binary: clean passthrough with missing env and with an unreachable allocator.