fix(ssd): reuse only finished prefetches; SIGTERM mid-stream test; doc fixes

[solderzzc]

Follow-ups from the SwiftLM review session's reviews of #196 and #197 (the owner approved them as one follow-up PR).

Changes

--stream-experts prefetch / load (from the #196 review)

1. Reuse only a finished prefetch. Because the streamed model now loads from its directory, the loader no longer fills in small missing files. A download interrupted after the last shard but before tokenizer.json or chat_template.jinja would therefore fail (or lose its chat template) on every later start. The reuse check now requires a .swiftlm-snapshot-complete marker. The marker is written only after snapshot returns and the copy has every shard plus the tokenizer files; otherwise the prefetch runs again.
2. Load from the directory only after the MoE check, and only with tokenizer files present. Before, the switch to ModelConfiguration(directory:) ran before the non-MoE check turned streaming off, so dense models started with --stream-experts were affected too. Copies without tokenizer files keep the id-based load, which can fetch them.
3. Re-validate after the snapshot. Offline, snapshot returns the repo directory even when it's partial. The prefetch now throws ModelDownloadIncomplete, reported as model_load_failed with a clear message, instead of activating streaming on a partial copy.
4. Last progress frame. finish() cancelled the Task before its final frame, so the bar stopped around 90%. It now draws one frame at the real fractionCompleted before ending the line. The frame rendering moved into ProgressTracker.frame(fraction:).
5. README: the crash warning gives the broadcast_shapes shape as an example (with top-k 8) instead of a fixed (N,8,8,D).

#197 review (docs and test)
6. exitAfterShutdownRequest() and its doc comment sat between emitEvent's /// block and func emitEvent, so they took over emitEvent's documentation. The function now sits after emitEvent.
7. The SIGPIPE comment now names exitAfterShutdownRequest() instead of Darwin.exit(0).
8. Test 39 in tests/test-server.sh: a SIGTERM sent while a stream is generating must exit with status 0 and put exiting{reason:"requested"} on its own JSON line. It must also leave no crash report for that server's PID; the check matches "pid" : N in .ips files so it doesn't pick up reports from other processes. Before this, only idle servers were SIGTERMed and the exit status was never read.

Verification (Mac mini M6)

• Test 39, run standalone against real binaries: b782 (exit(), before fix: exit with _exit after a requested shutdown (low priority) #197) fails with status=139 (SIGSEGV) and 1 crash report. This branch passes twice in a row. Local runs used mlx-community/SmolLM-360M-Instruct-4bit; CI uses the script's default model.
• Prefetch, with SmolLM-135M present only in the Application Support hub:
  • The first start runs snapshot once, writes the marker, and the bar ends at 100%.
  • The second start reuses the copy without a prefetch.
  • After removing the marker and a shard, the next start fetches again.
  • The offline-partial throw path wasn't exercised: this HubApi ignores HF_HUB_OFFLINE, so the shard was simply re-downloaded.
• Regression check, Qwen3.6-35B-A3B --stream-experts at 548 tokens: 14.16 tok/s decode at 5.7 GB. Streaming activates, with no "without streaming" line.

AI usage: written by Claude Code (Claude Opus 5.5) in the M6 benchmarking session, with the repo owner's approval to open this PR. The findings came from the SwiftLM review session's reviews.

🤖 Generated with Claude Code