v0.2.0 — measurement and distro awareness

Measurement and distro awareness: a held-out benchmark set, a benchmark of the gate itself,
per-model gate acceptance, and backport status from the distributions carried into the guide
and enforced by two new gate rules.

Added

• Backport status (bcpg.match.distro, bcpg scan --distro R --pkg-version V): each
  upstream CVE is open / fixed only in Ubuntu Pro / fixed by a backport / has no distro record,
  from the Debian and Ubuntu records in OSV for the installed package version. Shown in
  report.md (section and column), report.json and the trace. Rules gain distro_package.
• Holdout benchmark set: 33 libraries from Debian 12, Ubuntu 22.04 and 24.04, extracted
  after the rules were frozen (extract.py --set holdout). Frozen rules: 27/33 exact versions;
  after two rule fixes 29/33; the other four files carry no version string.
• bcpg gate-bench: false rejections of correct advice (0/540, 18 phrasings) and detection
  of 22 injected fault kinds (572/572), plus two out-of-scope semantic faults (0/54, reported).
• bcpg model-bench: per-model gate acceptance per batch, rejections by rule letter,
  tokens and time; results merge by model in bench/model_results.md.
• make fuzz: libFuzzer + ASan/UBSan harness for the C ELF parser in Docker
  (10 minutes, 2.38M inputs, no crash).
• README.ko.md and docs/design.ko.md (public design-decision record).
• docs/demo.gif (recorded with vhs docs/demo.tape, offline, no LLM).
• Distro-aware guide: with --distro, each CVE's backport status goes into the facts (with
  a per-CVE must_say line), the prompt and the template; validator rule (k) rejects an
  action for a CVE the installed package already fixes unless it says so, and a Pro-only fix
  unless it says Pro. Package and release versions (~18.04.23, Ubuntu 20.04) are no longer
  read as invented component versions. Sample: docs/sample/zlib-1.2.11-debian10/.
• Model benchmark row for qwen2.5 7B (4/9 batches accepted vs 6/9 for exaone3.5 2.4B).
• make lint; CI also runs ruff format --check.
• Validator rule (l): input field names (facts.…, allowed_refs, …) in the report text;
  rule (k) also rejects a "not fixed yet" phrase next to a CVE the distribution fixes or has no
  record of. Both from a qwen2.5 7B guide that passed the earlier rules.

Changed

• Offline OSV misses are logged as one line per run instead of one per CVE (47 lines for a
  holdout libcurl scan before).

Fixed

• report.md, report.json and trace.json are written with LF on Windows too.
• Validator rule (i) rejected correct advice that listed protocol versions ("TLS 1.2/1.3");
  found by gate-bench.
• Rules: libxml2's integer version glued to preceding bytes (::20914, Debian 12); zlib builds
  without copyright banners (bare ZLIB_VERSION, trusted inside libz.so only).

Key numbers: holdout set 29/33 exact versions (27/33 with the frozen rules) · gate 0/540 false rejections, 572/572 targeted faults detected · local LLMs on CPU: exaone3.5 2.4B 6/9 batches accepted, qwen2.5 7B 4/9.

Full changelog: CHANGELOG.md

v0.1.0 — first feature-complete release

First feature-complete release.

Pipeline

• extract — ELF parser written in C (elfdump.c, loaded with ctypes) with a pyelftools
  fallback that must produce identical results; PE via pefile (imports, VS_VERSION_INFO).
  Version-string candidates keep their file offsets. Corrupted inputs are bounds-checked and
  fuzzed in tests.
• identify — 10 CPE-grounded YAML signature rules (OpenSSL, zlib, curl, SQLite, Expat,
  libpng, BusyBox, PCRE2, libxml2, nginx); components are marked bundled (code inside the
  file, matched) or linked (a dependency to scan separately).
• identify --llm-hint (opt-in) — version strings no rule claimed go to the LLM; a
  suggestion is kept only if the version occurs verbatim at the cited offset and its CPE
  vendor:product is on an allowlist (every extra entry checked against the NVD CPE API on
  2026-10-23). Accepted/rejected counts are traced.
• match — NVD CPE query with a local range re-check, fixed_in from the range ends, OSV
  references as enrichment; disk cache and recorded fixtures for offline runs.
• retrieve — allowlisted advisory fetcher (redirects re-checked, size cap, per-host
  spacing), SQLite FTS5/BM25 index with a pure-Python BM25 fallback. Debian/Ubuntu tracker
  pages and recorded fixture pages are looked up for every CVE, not only NVD references;
  one passage per page is taken before a second from the same page.
• advise — Anthropic (structured output) or Ollama (JSON schema) writes the Korean guide
  from a facts JSON plus <document>-wrapped passages; injection-looking passages are flagged.
  Batch size is provider-aware (Ollama 4 CVEs per call, Anthropic 12; --batch-size
  overrides) after CPU Ollama timed out on 12-CVE batches; the Ollama request timeout is
  BCPG_OLLAMA_TIMEOUT (default 600 s). Null-like fixed_in strings from the model
  ("null", "None", "N/A", "없음", "", …) are coerced to null and whitespace is
  stripped before the gate; the count is traced as normalized_fields.
• validate — deterministic gate, rules (a)–(g) on the structured fields and (h)–(j) on the
  free text: inline citations must belong to the CVE (and to the finding's evidence), every
  version in the text must be installed / a fixed_in / a range bound, no upgrade target for a
  CVE without a fixed version, and priority words next to a CVE id must match its CVSS
  priority. Rejections are sent back verbatim (≤ 2 retries), then the template is used.
• report — Korean report.md, report.json, and trace.json with every stage, attempt,
  rejection reason, token count and page-fetch outcome.

Tools

• bcpg scan, bcpg identify, bcpg bench [--sets] [--online].
• bcpg-mcp MCP server: scan_binary, lookup_cve, identify_strings.
• Benchmark: synthetic (10 binaries) and distro (44 real libraries from Debian 10/11 and
  Ubuntu 18.04/20.04) — bundled-component precision/recall and version accuracy 100%.

Quality

• Test suite runs with zero network and zero LLM calls (recorded NVD/OSV/page fixtures,
  scripted fake provider); tests/conftest.py blocks non-loopback sockets and the Ollama
  port, stubs the Ollama probe and clears provider/API-key/offline variables, so local
  services and shell settings cannot change results. The Anthropic SDK is imported only when
  a real client is built or an SDK error is mapped, which took a cold-venv run from 46 s to
  13 s on Windows (about 10 s warm). CI on Ubuntu with the optional extras installed.
• Sample reports in docs/sample/: zlib 1.2.11 regenerated live with 0.1.0 (CPU Ollama; the
  gate rejected all three replies, so the template report shipped) and a template OpenSSL
  1.1.1n report.