docs: correct README record counts, schema version, links, competitive claims

[chaksaray]

Refs #79

Verification (re-measured 2026-10-11)

published records: 87   schema 1.1.0 on all 87   (dist/ave-records-latest.json)
covered records:   43   42 by pattern/YARA/Semgrep + AVE-2026-00024 by Magika
metadata only:      8   in scanner/ave_meta.py, no rule (00041-00045, 00049-00051)
coverage gap:      44   27 static_detection + 17 runtime-stage
toxic flows:       12   FLOW_DEFINITIONS in scanner/core/toxic_flows/flows.py
rules:             pattern 40, YARA 39, Semgrep 41

Script and record-by-record list: #79 (comment)

The first version of this PR said 80 published and 51 covered. The 80 went stale as the corpus grew. The 51 was wrong from the start: it counted records with metadata in scanner/ave_meta.py, not records with a rule, and the grep over scanner/rules/ it came from misses the pattern engine.

Corrected

• Record counts: published (87) and covered (43) are stated separately, along with the 8 metadata-only records and the 44-record gap. The 27 static and 17 runtime records are explained separately, because static analysis cannot detect the runtime-stage records by design.
• Tagline: removed "The only open-source scanner that produces OWASP AIVSS scores" (no citation) and "Never executes code", which the optional Docker sandbox stage contradicts. It now says Bawbel never starts MCP servers and runs no scanned content unless the sandbox is enabled.
• Toxic flows: removed "only open-source scanner that detects toxic flows". NVIDIA/SkillSpector ships TT1-TT5 taint chains plus AST8. Replaced with a claim about the mechanism.
• Comparison table: rebuilt with a source for every non-? cell (below). The execution row is split into "starts MCP servers" and "runs scanned skill content". The telemetry row, where SkillSpector leads, links to README accuracy: AVE coverage gap, telemetry, sync tooling, doc drift #79.
• Schema and links: schema is now 1.1.0. AVE links moved to aveproject.org and github.com/aveproject/ave. The dead api.piranha.bawbel.io link in Related is replaced with api.aveproject.org (returns 200 for 00001 through 00087).
• CI snippet: bawbel/scanner@v1 points at a tag that does not exist. It now pins @v1.3.0, whose tag includes action.yml. The snippet also passed an output: input the action does not declare, so the SARIF upload pointed at a file that was never written. It now uploads steps.bawbel.outputs.sarif-file.
• Pre-commit snippet: the documented hook (repo: bawbel/scanner, id: bawbel-scan) could never install, because this repo has never shipped .pre-commit-hooks.yaml. It's replaced with a repo: local hook that runs python -m scanner.bawbel_pre_commit. Tested in a scratch repo against bawbel-scanner==1.3.0 from PyPI: a clean file passes, and a copy of the golden fixture blocks the commit with exit 1.
• Docker: the images on Docker Hub are 1.2.3 (latest is the same digest as 1.2.3). The README claimed 1.3.0 and 1.3.0-full, which do not exist.
• Pipeline: added the missing Magika stage (Stage 0) to both diagrams, so "six engines" matches the code. "40+ regex rules" is now exactly 40 (40 distinct rule_ids).
• PROJECT_STRUCTURE.md: the action is in this repo, referenced as @v1.3.0. There is no .pre-commit-hooks.yaml here, and the hook definitions in bawbel/integrations are on main only, with no tag.
• Em dashes: 0 left in README.md.

Left unchanged

• Sample output still shows api.piranha.bawbel.io: that is the value every engine emits as piranha_url today. The host is now NXDOMAIN, so this is a code bug, not a doc question. The fix changes the value only, never the field name, and updates the README samples in the same PR. It is noted on README accuracy: AVE coverage gap, telemetry, sync tooling, doc drift #79.
• scanner/ave_meta.py docstring still cites schema 1.0.0. It is a comment, tracked in README accuracy: AVE coverage gap, telemetry, sync tooling, doc drift #79.
• PIRANHA_API_URL vs BAWBEL_PIRANHA_URL: config and docs disagree. Tracked in README accuracy: AVE coverage gap, telemetry, sync tooling, doc drift #79.

Sources for every competitive-table cell

• Bawbel: this repo's own code/docs.
• NVIDIA SkillSpector: https://github.com/NVIDIA/SkillSpector README, fetched
  2026-09-12 (execution model, taint chains TT1-TT5 + AST8, baseline suppression
  glob+fingerprint no expiry, metadata.inference_usage, SARIF 2.1.0, Apache 2.0,
  no AIVSS/CVSS, no MCP conformance grading, no rug-pull detection, internal rule IDs
  with OSV.dev integration for CVEs).
• Snyk agent-scan: https://github.com/snyk/agent-scan README, fetched 2026-09-12
  (explicitly executes MCP stdio servers, "Toxic Flows" detection category present but
  chain count not published, --ignore-issues-codes/--ignore-risks suppression with
  no expiry/audit trail, no SARIF, no MCP conformance grading, Apache-2.0).
• Cisco DefenseClaw: https://github.com/cisco-ai-defense/defenseclaw README,
  fetched 2026-09-12. Only License (Apache-2.0) was clearly confirmable from the
  excerpt; every other cell is marked ? rather than asserting an unconfirmed
  capability. License re-checked 2026-10-11 via the GitHub API: Apache-2.0.

Not in scope for this PR

• No changes under scanner/, config/, or tests/.
• Lint & Format previously failed here because of the missing trailing newlines on the base branch. Fixed by fix: add missing trailing newlines to unblock CI lint #83 and fix: restore trailing newlines stripped by 9292c2c #85; develop is merged into this branch and all checks pass.

🤖 Generated with Claude Code