bug(score): grounding checker counts URLs, env vars, and HTTP paths as broken file references

[stackbilt-admin]

Problem

charter score on aegis-daemon reports 177 broken path references. The top entries reveal the path resolver is not filtering non-file strings:

• AEGIS_LOCAL_URL=http:/localhost:11434 — env var assignment pulled from a code block
• C:/Users/kover/Documents/aegis-daemon — Windows absolute path in a markdown table
• /chat, /api/runs, /webhooks/voice/complete — HTTP route paths in a table

None of these are file paths, but the grounding checker treats any string that looks like a path as a reference to validate.

False positive categories

1. URL strings — anything containing :// or starting with http should be excluded
2. Env var assignments — KEY=value lines in code blocks should not have the value extracted as a path
3. HTTP routes — strings starting with /api/, /webhooks/, /chat etc. in markdown tables are URL paths, not filesystem paths
4. Windows absolute paths — C:/... paths in documentation tables are reference documentation, not resolvable local paths
5. Env var names — bare ANTHROPIC_API_KEY tokens are not file paths

Impact

These false positives inflate the broken reference count from a likely real count of ~5–10 to 177, dragging the Grounding score from what should be ~18/20 to 7/20. This makes the score misleading and undermines trust in the metric.

Fix

Before running path resolution:

1. Strip any token that matches https?:// (URL)
2. Strip any token that matches ^[A-Z_]+= (env var assignment)
3. Strip any token that starts with /api/, /webhooks/, /health, /chat (HTTP route heuristic, or: only validate paths that resolve to a plausible relative file under the repo root)
4. Strip Windows absolute paths ([A-Z]:/)
5. Only validate paths containing a file extension (.ts, .md, .json, etc.) OR paths that start with ./ or ../ (explicit relative references)

Alternatively: require file references to appear in a specific ADF notation (e.g., ref:path/to/file) rather than doing heuristic path extraction from all text.

[stackbilt-admin]

Fixed in PR #168