ARCHITECTURE.md

@@ -0,0 +1,79 @@
+# Architecture — live.f12.rocks
+
+## Overview
+
+```
+Phone/Syncthing                Docker host
+───────────────                ───────────
+incoming/  ──watch──►  worker  ──►  variants/ + intermediates/
+                              │         │
+                              └── rembg + gmic pipeline
+                                        │
+meta/{stem}.json ◄──────────────────────┘
+       │
+       ▼
+     web (FastAPI)  ──►  browser UI + remix POST
+```
+
+Two containers share one image (`Dockerfile`) and one bind-mounted data
+directory (`DATA_HOST_DIR` → `/data`). Syncthing runs on the host, not in
+compose.
+
+## Python modules
+
+| Module | Role |
+|--------|------|
+| `app/worker.py` | Poll `incoming/`, copy to `variants/`, run `process_job()` |
+| `app/pipeline.py` | gmic/rembg orchestration, meta/manifest I/O, variant compose |
+| `app/remix.py` | User-selected remix variant (one-shot compose + manifest append) |
+| `app/main.py` | FastAPI routes, Jinja templates, static files |
+| `app/config.py` | Environment → constants (paths, timeouts, tuning) |
+| `app/rembg_cli.py` | Subprocess entry for rembg (avoids heavy `rembg[cli]` extra) |
+| `app/io_utils.py` | Atomic JSON read/write |
+| `app/env_utils.py` | Shared bool env parsing |
+| `app/logging_config.py` | Entry-point logging setup |
+
+## Data model
+
+One **job** = one sanitized stem (from incoming filename). Flat files:
+
+- `variants/{stem}_original.jpg` — preprocessed copy
+- `variants/{stem}_rembg.png` — background removal
+- `variants/{stem}_vN.png` — final composed images
+- `intermediates/{stem}_vN_*.png` — step images (kept for debugging)
+- `meta/{stem}.json` — status + manifest (variants list, filter metadata)
+- `processed.json` — worker bookkeeping (incoming path → handled)
+
+## Pipeline origin
+
+The gmic filter/blend chain is ported from an external reference script
+`make_random.py` (not in this repo). `config.POST_FILTERS` and asset lists
+under `assets/` mirror that script's behaviour.
+
+## Background removal choice
+
+`compare-bg/` is a **standalone benchmark** (rembg models vs withoutbg).
+It is not part of the runtime stack. Production uses `u2net` + alpha
+matting via `app/rembg_cli.py` (see `config.REMBG_MODEL`).
+
+## Known limitations (intentional)
+
+- **No web auth** — anyone with the link sees all jobs (event tool).
+- **Sequential worker** — one photo at a time; burst uploads queue.
+- **Meta JSON** — worker and web both read-modify-write `meta/*.json`
+  without file locking; concurrent remix during active processing can race.
+- **Incoming is append-only** — worker never deletes from `incoming/`.
+
+## Dev / test
+
+```bash
+python3.11 -m venv .venv && source .venv/bin/activate
+pip install -r requirements-dev.txt
+ruff check app/ tests/
+pytest -q
+```
+
+Full stack (gmic/rembg): `docker compose up --build` with `DATA_HOST_DIR=./data`.
+
+Python **3.11** matches the Docker image; host 3.14+ cannot install pinned
+`onnxruntime` — use Docker or a 3.11 venv for parity.

CLEANUP_REPORT.md

@@ -0,0 +1,89 @@
+# Cleanup Report — live.f12.rocks
+
+Branch: `cleanup/code-quality` (local only, not pushed).  
+Completed: 2026-07-18.
+
+## Summary
+
+Incremental code-quality cleanup across 10 commits. **No intentional
+behavior changes** to the photo pipeline, worker loop, or web UX.
+
+## What changed
+
+| Step | What | Why | Risk | Verification |
+|------|------|-----|------|--------------|
+| Phase 0 | `CLEANUP_PLAN.md` | Analysis + roadmap | Low | — |
+| Tooling | `pyproject.toml`, `requirements-dev.txt`, `.python-version` | Ruff + pytest baseline, pin 3.11 | Low | `ruff check`, `pytest` |
+| Style | Ruff format entire `app/` | Consistent style | Low | `ruff format --check` |
+| Dead code | Removed unused `blend_layers()` | Less noise | Low | grep + pytest |
+| `.gitignore` | Build/coverage/venv patterns | Keep repo clean | Low | — |
+| `io_utils` | Atomic JSON helpers | DRY meta/processed I/O | Low | `test_io_utils`, pipeline/worker tests |
+| `env_utils` / `logging_config` | Shared bool env + logging | DRY | Low | `test_env_utils`, pytest |
+| `compose.yml` | YAML anchor for env | DRY, fewer drift bugs | Low | `docker compose config` (manual) |
+| Config | `PREPROCESS_TIMEOUT` constant | Named magic number | Low | default still 120s |
+| `read_meta()` | Public meta accessor; `list_jobs` reads once | Perf + clarity | Low | `test_pipeline`, `test_main` |
+| Tests | 33 characterization tests | Safe refactor baseline | Low | `pytest -q` |
+| CI | `.gitea/workflows/ci.yml` | Lint + test on push/PR | Low | workflow syntax |
+| Docs | `README.md`, `ARCHITECTURE.md` | Onboarding | Low | — |
+
+## What was NOT changed (and why)
+
+| Item | Reason |
+|------|--------|
+| `pipeline.py` module split | Medium risk without broader integration tests; deferred |
+| Meta file locking | Behavior change; documented in `ARCHITECTURE.md` |
+| Dependency major bumps | Per plan — list only, no auto-bump |
+| `compare-bg/` | Separate experiment; left untouched |
+| Deploy CI | Manual deploy on boka stays; STANDARDS says ask first |
+| FastAPI `TemplateResponse` API migration | Deprecation warning only; no functional change |
+| `_validate_job_id` always sanitizes to valid ID | Pre-existing; documented in tests |
+
+## Bugs noticed (not fixed)
+
+1. **Meta JSON race** — worker and web concurrent RMW without locking.
+2. **`_filter_timeout_override` global** — thread-unsafe under concurrent remix (uvicorn workers).
+3. **`sanitize_stem("a/b")` → `"b"`** — `Path.stem` drops path prefix before slash replacement.
+4. **`_validate_job_id`** — sanitization makes rejection path effectively unreachable.
+
+## Dependency notes (suggestions only)
+
+| Package | Current pin | Note |
+|---------|-------------|------|
+| `fastapi` | `<0.117` | 0.117+ available; test before bump |
+| `onnxruntime` | `<1.20` | Very old pin; rembg may allow newer in future |
+| `pillow` | `<11` | Pillow 11 exists; verify rembg compatibility |
+
+No lockfile added — soft pins kept to avoid changing Docker resolve behavior.
+
+## How to verify locally
+
+```bash
+git checkout cleanup/code-quality
+python3.11 -m venv .venv && source .venv/bin/activate
+pip install -r requirements-dev.txt
+ruff check app/ tests/
+pytest -q
+# Full stack:
+cp .env.example .env   # set DATA_HOST_DIR=./data
+docker compose up --build
+```
+
+## Commits on branch
+
+```
+docs: add cleanup plan from phase-0 analysis
+chore: add pyproject.toml with ruff and pytest config
+style: apply ruff format and extend gitignore
+refactor: extract atomic JSON helpers to io_utils
+refactor: share env-bool parsing and logging setup
+chore: deduplicate compose environment blocks
+refactor: named preprocess timeout and single meta read
+test: add characterization tests for core modules  (+ CI workflow)
+docs: update README and add ARCHITECTURE
+```
+
+## External references
+
+- `compare-bg/` — local venv `.venv-compare/` and 1.5GB model cache (gitignored)
+- `make_random.py` — external reference script, not vendored
+- Submodule/symlink: none

README.md

@@ -55,6 +55,24 @@ Drop a photo into `incoming/` (via Syncthing or locally). After a few
 seconds (poll interval + processing time) it shows up on the web UI; flat
 files land in `variants/` and `intermediates/` for the phone.
 
+## Development
+
+Python **3.11** (same as `Dockerfile`). On a host with newer Python, use
+Docker or a 3.11 venv — pinned `onnxruntime` does not publish wheels for 3.14+.
+
+```bash
+python3.11 -m venv .venv && source .venv/bin/activate
+pip install -r requirements-dev.txt
+
+ruff check app/ tests/       # lint
+ruff format app/ tests/      # format
+pytest -q                    # unit tests (no gmic/rembg)
+```
+
+CI (`.gitea/workflows/ci.yml`): ruff + pytest on push/PR.
+
+Architecture notes: `ARCHITECTURE.md`. Cleanup log: `CLEANUP_PLAN.md`.
+
 ## Configuration (`.env`, see `.env.example`)
 
 | Var | Default | Notes |
@@ -64,6 +82,8 @@ files land in `variants/` and `intermediates/` for the phone.
 | `BLEND_OPACITY` | `30%` | Default blend opacity for auto-generated variants |
 | `FILTER_TIMEOUT` | `120` | Seconds before a single gmic call is killed |
 | `MAX_FILTER_ATTEMPTS` | `8` | Random-filter retry budget per bg/fg pick |
+| `MAX_EDGE_PX` | `2000` | Longest edge after ImageMagick preprocess |
+| `PREPROCESS_TIMEOUT` | `120` | Seconds for preprocess step |
 | `NICE_LEVEL` | `18` | `nice -n` level for gmic/rembg subprocesses |
 
 ## Deploy notes / caveats