feat: v0.9.0 — cf-text, pipeline crystallization engine, multimodal pipeline, a11y preferences
Closes #33, #37, #38, #41, #42.
## cf-text (closes #41)
- New module: `circuitforge_core.text` — direct local inference bypassing ollama/vllm
- Backends: llama.cpp (GGUF), transformers (HF), mock
- Auto-detects backend from file extension; CF_TEXT_BACKEND env override
- Optional 4-bit/8-bit quantisation via bitsandbytes (CF_TEXT_4BIT / CF_TEXT_8BIT)
- process-level singleton + per-request `make_backend()` path
## Pipeline crystallization engine (closes #33, #37)
- FPGA→ASIC model: LLM-discovered paths → deterministic workflows after N approvals
- `models.py`: PipelineRun (incl. review_duration_ms + output_modified per #37),
CrystallizedWorkflow, Step, hash_input()
- `recorder.py`: append-only JSON run log under ~/.config/circuitforge/pipeline/
- `crystallizer.py`: threshold check, majority/most-recent step strategy,
rubber-stamp warning (review_duration_ms < 5s triggers warnings.warn)
- `registry.py`: exact + fuzzy match, deactivate-without-delete, colon-safe filenames
- `executor.py`: deterministic steps with transparent LLM fallback
## Multimodal chunked pipeline (closes #42)
- `pipeline/multimodal.py`: cf-docuvision pages → cf-text streaming
- `run()` yields PageResult per page (progressive, no full-doc buffer)
- `stream()` yields (page_idx, token) tuples for token-level UI rendering
- `vram_serialise` flag + `swap_fn` hook for 8GB GPU VRAM management
- `prompt_fn` callback for product-specific prompt construction
## Accessibility preferences (closes #38)
- `preferences/accessibility.py`: PREF_REDUCED_MOTION, PREF_HIGH_CONTRAST,
PREF_FONT_SIZE, PREF_SCREEN_READER with get/set helpers
- Exported from preferences package __init__
## LLM router fix
- cf-orch backends: skip reachability pre-check; allocation starts the service
- Static backends: reachability check remains in place
pyr0ball
80b0d5fd34