serve: viz scans entire project root — ignores flows_dir, .gitignore, and non-flow YAML

[nullhack]

Summary

flowr serve discovers every *.yaml file under the project root and treats each as a flow. Discovery ignores [tool.flowr].flows_dir, ignores .gitignore, and has no exclude mechanism, so archived/copied flow files and non-flow YAML anywhere in the tree show up in the viz alongside the real flows (and non-flow YAML either renders as a bogus "invalid" flow or fails to parse in the UI).

The CLI commands already scope themselves to flows_dir (default .flowr/flows); only the viz server scans the whole root.

Steps to Reproduce

mkdir -p /tmp/flowr-scan-demo && cd /tmp/flowr-scan-demo
mkdir -p .flowr/flows archive config

# one real flow
cat > .flowr/flows/main.yaml <<'EOF'
flow: main
version: "1.0.0"
states:
  - id: start
    next:
      done: done
  - id: done
EOF

# an archived copy of a flow, elsewhere in the tree (gitignored in real use)
cp .flowr/flows/main.yaml archive/old-main.yaml

# an unrelated YAML config file
cat > config/app.yaml <<'EOF'
runtime: python
region: us-east-1
EOF

flowr serve --path .

Observed

The viz lists three flows: main, old-main, and config/app. config/app.yaml has no flow/states keys, so it's indistinguishable from a real flow in the list and surfaces as a parse error when opened.

Expected

Only main is discovered. Discovery should be bounded by flows_dir (matching the CLI) and/or exclude .gitignore'd paths and files that don't structurally look like flows.

Root Cause

flowr/server/scanner.py, FlowRegistry._refresh():

def _refresh(self) -> None:
    self._files = []
    for yf in sorted(self._root.rglob("*.yaml")):
        rp = yf.relative_to(self._root)
        self._files.append(FlowFile(name=yf.stem, relative_path=str(rp)))

The registry is constructed with the raw --path value (. by default) from cli/serve.py:cmd_serve, and discover_flows(path) passes it straight through. Nothing in the scan path consults:

• FlowrConfig.flows_dir / [tool.flowr].flows_dir,
• .gitignore, or
• the existing _check_structure() validator (which already requires flow + states and is only used by write_flow/create_flow, not by discovery).

Proposed Fix

Pick one (or combine):

1. Scope discovery to flows_dir — read [tool.flowr].flows_dir (default .flowr/flows) and rglob only under it. This aligns the viz with the CLI's existing behaviour and is the smallest change. Falls back to . only if flows_dir is unset.
2. Honor .gitignore — skip paths matched by the project's .gitignore (e.g. via pathspec). Handles the common "archived/backup flow copies" case without config.
3. Filter by structure at discovery time — reuse _check_structure() to drop any YAML lacking flow + states before adding it to the registry. Cheap defense-in-depth that also rules out unrelated config YAML.

Option 1 alone fixes the reported case; options 2–3 make it robust against edge layouts.

Impact

• Any project that keeps non-flow YAML (config, compose files, schema, vendored docs) or archived flow copies anywhere under the root gets a polluted, misleading flow list in the viz.
• Users have no workaround except moving those files out of the project tree or renaming their extensions, because serve exposes no --flows-dir / --ignore flag and --path doubles as the scan root.