fix: harden impact LLM JSON parsing

[AusafMo]

Summary

• increase the impact LLM response budget so richer narrated overviews are less likely to truncate
• tighten the provider prompt to request terse raw JSON without Markdown fences
• recover useful top-level impact fields when a provider returns truncated/partial JSON
• update installation docs to use the published PyPI packages: cfgit and cfgit-impact
• remove internal project notes and the obsolete archived v0.1 spec from the public tree
• ignore local project notes and recording/video drafts so they do not get committed accidentally

Public/PyPI scrub

• removed docs/project-notes/* from git
• removed docs/archive/spec-v0.1.md because the current SPEC.md / SPEC_CORE.md are the public references
• checked built sdists/wheels for project notes, archive docs, and video files: none found

Tests

• .venv/bin/python -m pytest tests/test_impact_boundary.py -q twice: 12 passed
• .venv/bin/python -m pytest -q twice before docs cleanup: 73 passed, 2 skipped
• .venv/bin/python -m pytest -q twice after docs cleanup: 73 passed, 2 skipped
• .venv/bin/python -m build: passed for cfgit
• cd plugins/cfg_impact && ../../.venv/bin/python -m build: passed for cfgit-impact
• clean wheel install smoke: cfg --help, cfg-mcp --help, import cfg, import cfg.mcp.server, import cfg_impact passed
• artifact scan: no project-notes, archive, or video files in cfgit/cfgit-impact sdist or wheel
• git diff --check: passed