fact-gate 1.2.0 — git pre-commit mode, sticky PR comment

Where it runs, part two: git and the PR conversation. Chosen after a second
round of adversarial review of the roadmap, which dropped an MCP server and
a browser playground (no evidence either channel produces users for a CI
tool) in favour of the two things every team already has.

Added

• fact-gate pre-commit. Checks the staged content (git show :path)
  of every file the nearest .fact-gate.json covers, refuses the commit on
  a contradiction. Editor-agnostic: Cursor, Codex, Copilot and humans go
  through the same gate. .pre-commit-hooks.yaml for the pre-commit
  framework; husky and lefthook one-liners in the README. checkStaged,
  stagedPaths, stagedContent in the library.
• Sticky PR comment. comment: 'true' on the Action posts one comment
  with the report and keeps it updated on every push (pull-requests: write).
  The Action now records its exit code and fails in a final step, so the
  comment is posted even when the gate blocks.
• fact-gate/core: the browser-safe half of the library (no Node built-ins).

Deferred, deliberately

MCP server, GitHub Pages playground, SARIF output, dedicated JUnit / lcov /
git-log adapters. The cmd measurement already covers any artefact a shell
can read; the rest will be built when a user asks.

fact-gate 1.1.0 — PR descriptions vs the diff, measured facts, GitHub Action, Claude Code hook

The gate moves from "run it by hand on a CV" to "it runs where the text is
written". Same engine, three new places it fires, and three new kinds of
truth it can check against.

Added

• A diff is a source of truth. --diff <file> / --git <base> /
  verifyFacts(text, { diff }) turn a unified diff (or git diff --numstat)
  into facts: files by status, lines by direction, test cases added,
  dependencies added or removed (npm, pip, go.mod, Cargo). A PR description
  is verified against the change it describes. parseDiff, factsFromDiff.
• Measured facts. Any value in a facts file may be {"cmd"|"file"|"json", "pattern"} and is resolved at check time, so the README is checked against
  the test suite's own output, not against a number someone typed last month.
  fact-gate measure, resolveFacts, measure. A measurement that fails is
  an error, never a silent skip.
• GitHub Action. uses: AnonZ7/fact-gate@v1 checks the PR body against
  the PR diff (default) or any file against facts/source; annotations, a
  step summary with the measured facts, and verdict / report outputs.
• Claude Code hook. fact-gate hook as a PreToolUse hook on Write / Edit
  / MultiEdit denies a save whose numbers contradict .fact-gate.json and
  hands the reasons back to the agent. fact-gate init scaffolds the config
  and prints the settings snippet. Inert when no config is found; never
  breaks a session.
• Extraction: verbs qualify a count (adds 4 tests → tests/added);
  participles after the noun (6 lines deleted); floors from words (over 4,000, more than 30, at least 5); approximate claims (roughly 500,
  about 5k, ~85) judged with a 10% tolerance; spelled-out small counts
  (two new files) and no in a change context (no new dependencies = 0
  added); colon breakdowns (6 files: 2 new, 1 deleted, 3 modified);
  qualified noun-first forms (hook files: 27); singular nouns (1 dependency); a code-change vocabulary (files, insertions, deletions,
  dependencies, packages, functions, …).
• Comparison: a qualified claim is judged by the source entries that share
  its qualifier before the generous number-only match applies, so no new dependencies meets dependencies added: 1 and loses instead of meeting
  dependencies deleted: 0 and winning. Contradiction messages name the
  best-matching figure with its qualifiers.
• Output: --format markdown (table for PR comments), --format github
  (annotations + $GITHUB_STEP_SUMMARY + $GITHUB_OUTPUT), colour on a TTY,
  --target-env, --no-measure, --no-color.
• Evals: three suites (CV vs profile, PR body vs diff, docs vs facts)
  with per-suite precision and recall, so a weak domain cannot hide in the
  average. evals/run.mjs --json.
• JSON Schema for .fact-gate.json; mergeFacts; TypeScript declarations
  for everything above.
• The repository gates its own README with itself
  (.github/workflows/fact-gate.yml, .fact-gate.json).
• Trust model for command measurements. In hook mode a config's cmd
  measurements run only after fact-gate trust <config> has recorded the
  file's hash in ~/.fact-gate/trusted.json (or with FACT_GATE_ALLOW_CMD=1
  in CI). Until then literal, file and json facts still gate, and the
  agent is told why the command was skipped. Editing the config revokes trust.
• Negated counts (does not touch 12 files, doesn't add 40 tests,
  without adding 300 lines) are not claims of that count.
• Fullwidth and Arabic-Indic digits normalise to ASCII before extraction.
• Unknown --flags are a usage error (exit 2) instead of being ignored.

Fixed

• Hook ordering (pre-release review, critical): the include/exclude filter
  now runs before anything in the config is loaded, so a write to a file the
  config does not cover runs nothing from the config. Pinned by
  test/security.test.mjs.
• A heading on the line above a Tests: 999 claim was captured as a
  qualifier, turning a fabrication into a warning. Noun-first qualifiers must
  now sit on the same line; line breaks survive normalisation.
• An Edit whose old_string is absent from the file is skipped instead of
  being judged against the unmodified file.
• A currency amount followed by a noun ($90,000 deal) was also read as a
  count once singular nouns entered the vocabulary. Counts now refuse a
  currency prefix.
• total / overall / combined are vacuous qualifiers and no longer make
  total tests: 900 incomparable with tests: 846.

Changed

• contradicts() is kept; contradictions() returns the ranked entries the
  messages are built from.