fix: make the 512gb tier's glm-5.2 actually serve

[jerrytalton]

Problem

v1.2.0's 512gb tier points general/reasoning/long_context/translation at glm-5.2, but the model had never loaded on the 512GB Studio: mlx-lm (all released versions and main) can't load GLM-5.2's shared-indexer layout (config.indexer_types) and fails strict load with "Missing 285 parameters". The release gate didn't catch it because the correctness tests skip when a model isn't pulled — and the release was cut on a machine that can't hold the 390GB checkpoint. On top of that, two mlx-openai-server bugs (still present in 1.8.1) meant even a loadable 390GB on-demand model couldn't stay resident: a hardcoded 300s readiness timeout, and a warm-request fast path that bypasses refcounting so the idle timer unloads the model every 300s regardless of traffic — thrash-reloading 390GB every ~5 minutes against the menu bar's 240s keep-warm pings.

Changes

• bin/apply-mlx-glm52-patch.sh — idempotent script that ports the two model files from unmerged upstream PR GLM-5.2: full/shared indexer typing for glm_moe_dsa (DSA schedule + interleaved indexer rope) ml-explore/mlx-lm#1463 onto the installed mlx-lm 0.31.1 (pinned to the PR head sha, fails closed if the ref moves; installing the PR branch wholesale is blocked by the mlx ≥0.31.2 stream-hang, mlx-lm#1256), raises the readiness timeout to 1800s, and fixes the warm-request refcount bypass. Self-disables once upstream ships indexer_types support. install.sh runs it automatically on 512GB machines.
• app/profile-server.py — chat read timeout 300→600s; a cold frontier-model load stacking with a thinking-model generation exceeded 300s on valid requests.
• tests/test_tools_smoke_*.py — tier model maps now derive from DEFAULT_PROFILES instead of hardcoded copies that drifted twice.
• docs — troubleshooting entry (symptoms, root causes, fix, single-on-demand-slot limitation and the whisper-static mitigation) and CLAUDE.md notes.

Verification (on the 512GB Studio)

• glm-5.2 strict-loads clean and answers correctly; PR #1463's unit tests pass against the ported files.
• Long-context probe at 5,108 prompt tokens (past the index_topk=2048 sparse-attention engagement point where mlx-lm#1453 reported gibberish): correct recall from start/middle/end.
• Warm-keeping verified: idle timer re-arms on every 240s keep-warm ping; no more unload/reload cycle.
• Full correctness suite: 13/13 (one image_gen Metal GPU-timeout flake reproduced only while a 390GB reload was in flight — the thrash this PR eliminates — and passed cleanly on retry).
• Patch script exercised through full path, fast path, and idempotent re-run; installed files byte-identical to the PR head.

🤖 Generated with Claude Code