)

* fix: update TemplateResponse calls for Starlette 1.0 compatibility

Starlette 1.0.0 changed the TemplateResponse signature from
TemplateResponse(name, context) to TemplateResponse(request, name, context).

The old positional API passed the context dict as the `name` parameter,
causing Jinja2's LRUCache to receive an unhashable dict as a cache key:

    TypeError: unhashable type: 'dict'

This broke the admin dashboard (login, dashboard, chat pages) with
HTTP 500, while the /v1/ inference API was unaffected.

- Update 3 TemplateResponse calls in admin/routes.py
- Pass explicit empty context dict for consistency
- Add TestLoginPage and TestDashboardPage test classes
- Update existing TestChatPageApiKeyInjection assertions

* address review: bump fastapi floor + improve login_page test assertion

- Bump fastapi>=0.100.0 to >=0.108.0 to match the Starlette 1.0
  TemplateResponse(request, name, context) signature requirement
  (Codex P1 review)
- Use assert_called_once_with in login_page test to verify all args
  including context dict (Gemini review)

…370)

animate-fade-in-up ended with transform: translateY(0) which creates
a stacking context per card, trapping popup z-50 inside. changed to
transform: none which doesn't create a stacking context.

…tring (#367)

…372)

- defer scheduling new requests when active memory exceeds soft limit
  during generation to prevent Metal allocation failures
- clear Metal buffer cache in fail_all_requests() to reclaim fragmented
  memory after batch generation errors
- rename "Prefill memory guard" to "Memory guard" in admin UI and move
  toggle to top of Resource Management section

detect adapter_config.json to filter out LoRA/PEFT adapters that oMLX
cannot load. show warning badge instead of download button in admin UI.

_IntOffsetCacheProxy returned _idx for RoPE offset, but
BatchRotatingKVCache._idx wraps at max_size (e.g. 1024 -> 0).
after the sliding window fills, RoPE positions reset to 0 causing
gibberish output on Gemma3 and other sliding-window VLM models.

use _offset (monotonic, never wraps) instead. only
BatchRotatingKVCache has _offset, so BatchKVCache and other cache
types are unaffected.

replace _idx/_offset shortcuts with direct offset[0].item() extraction.
_idx wraps at max_size (continuous generation), _offset diverges after
merge() which sets it to buffer size instead of actual token count
(SSD cache restore). the mx.array offset is always correct.

also add Gemma3-12B-QAT to boundary cache consistency tests.

- implement GPTQ column-wise error compensation for all quantizable weights
- batched expert GPTQ processes 256 experts simultaneously (15x faster)
- shared Hessian across experts in each MoE layer
- sensitivity budget assigns per-tensor bits before GPTQ optimization
- fix float32 norm weights to bfloat16 for mlx-lm inference parity
- add eos_token_id from generation_config.json
- add Step-3.5 MoE support (moe.*_proj pattern)
- update admin UI: Enhanced Quantization(+) with GPTQ description
- update docs/oQ_Quantization.md for GPTQ-based pipeline
- clean up legacy equalization code

codebook-quantized KV cache that reduces memory ~50-70% during decode
with near-lossless quality. lazy quantization keeps prefill at fp16
speed, then compresses at decode start.

core:
- TurboQuantKVCache + BatchTurboQuantKVCache with full batch lifecycle
- 2-pass fused Flash Attention Metal kernel (no dequant during decode)
- boundary-based quantization (19x faster than argmin)
- batch decode_attention via same fused kernel (B>1 grid dispatch)

integration:
- attention patch with VLM _IntOffsetCacheProxy unwrap
- prefix cache: save quantized blocks to SSD, dequant to KVCache on
  restore for merge compatibility. meta_state stores (offset, bits, seed)
- type_registry: TurboQuantKVCache recognized as sliceable
- admin UI: turboquant toggle with 3-bit/4-bit selection