fix(ssd): stream the directory we load, resume partial first-run downloads, keep exiting events on their own line

[solderzzc]

Follow-ups from the SwiftLM review session's review of #192 (the owner OK'd them as post-merge work), plus one gap found while testing them.

Changes

• Streaming uses the directory that's actually loaded. Since Fix SSD streaming buffer overflow with MTP, and scope streaming to its model mlx-swift-lm#71, loadWeights streams only when the loaded directory matches the one streaming was activated for. modelConfig and the streaming directory came from different lookups, so a model found only by resolveModelDirectory (for example ~/.cache) got a second copy downloaded into Application Support and loaded without streaming. When --stream-experts is set, modelConfig now points at modelDirectory.
• Found while testing: resolveModelDirectory doesn't check that the weights are present. An empty or partial ~/.cache snapshot (for example from an interrupted download) resolved as the model directory, so the prefetch was skipped and streaming was activated for a directory the loader would never read. With --stream-experts, a resolved directory that fails ModelStorage.validateLocalModelDirectory is now dropped.
• Partial first-run download: the prefetch reuse check used to look only for config.json. The architecture probe or an interrupted download can leave that behind, and the session then planned with a partial size. It now validates every shard, and otherwise resumes through snapshot.
• Progress bars: ProgressTracker.finish() stops redrawing and ends the \r line. It runs when the prefetch or the main load returns or throws. emitEvent also starts a fresh line if a bar is still open. Before, a mid-download failure glued the exiting JSON onto the progress line, where the Aegis daemon (serde_json per line) and the Electron manager (startsWith('{')) couldn't see it. This problem already existed for the loader's bar.
• README: the b769/b773 crash warning now covers quantized MoE models generally. It reproduced on M6 with Gemma 4 26B-A4B too: broadcast_shapes (260,8,8,2816) vs (260,8,1) on b769. The Gemma 4 8-bit --stream-experts row predated fix: propagate throws from LanguageModel.newCache and adapt to mlx-swift-lm PR #62 #167, so I re-measured it: 9.1 tok/s (was 8.8), 9.5K tokens ok, 32K swaps, 7.3 GB. The data is appended to docs/profiling/m6/gemma4_26b_a4b_8bit.*.

Verification (Mac mini M6, 32 GB)

• Interrupted download: run 1 was killed mid-download, leaving config.json plus the tokenizer files. Before the directory-validation fix, run 2 skipped the prefetch, because an empty ~/.cache snapshot resolved as the model directory. With it, the next start re-downloads, finishes the bar with a newline, prints the next log line on a fresh line, and reaches ready. Run 3 reuses the Application Support copy without downloading.
• Regression, Qwen3.6-35B-A3B --stream-experts, 548 tokens: 248.6 prefill / 13.89 decode tok/s at 5.69 GB, with no "loading … without streaming" line. Streaming still activates for a cached model.
• Gemma 4 26B-A4B 4-bit --stream-experts on the fixed build: 16.75 / 17.01 tok/s decode at 548 / 2.3K tokens, 4.4–5.1 GB.
• The emitEvent newline after a mid-download network failure wasn't reproduced; it's based on code reading.

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 review findings came from the SwiftLM review session.

🤖 Generated with Claude Code