feat(ci): native CI regression gate over registry verdicts

[HarperZ9]

What this ships

A native CI regression gate: crucible ci REGISTRY.

• --write-baseline FILE captures the current verified-latest verdict per (thesis, claim) as a sealed baseline snapshot (crucible.ci-baseline/v1).
• --baseline FILE re-derives the current verdicts, compares them against the baseline, exits nonzero on regression, and emits a deterministic PR-comment-ready Markdown claims-by-rounds verdict matrix.
• A regression is a claim moving MATCH -> DRIFT, becoming UNVERIFIABLE, or dropping out of the verified-latest set (fail-closed: a claim that lost its witnessed standing is never read as held). New claims and improvements do not fail the gate.
• Every summary cell references the re-derivable assessment packet (the assessment seal); a reviewer re-checks any cell with crucible verdicts REGISTRY --verify.
• --out FILE writes the Markdown while preserving the gate exit code; --json emits the full gate report.
• The baseline file seals its own cells, so a hand-edited baseline (a downgraded status) is rejected on load.

The feature rides on the existing accountability layer: it re-reads the same verified-latest state registry stats uses and opens no second verdict path. The gate's own output is a pure function of two snapshots, so it recomputes from the record and cannot be asserted.

Structure

• crucible/ci_gate.py (255 lines) - the pure core: Cell / Snapshot (sealed) / GateReport, status ranking, gate(), cells_from_latest().
• crucible/ci_report.py (67 lines) - deterministic Markdown rendering (split from logic per the repo's report.py vs assess.py convention, keeping every file under the 300-line budget).
• crucible/ci_cmd.py (85 lines) - the crucible ci command handler.
• cli.py / __init__.py - wiring and public API exports.

Docs

• README gains a CI regression gate section with the command and a ready-to-copy GitHub Action snippet (assess -> gate -> post the Markdown as a PR comment, if: always() on the comment step so the exit code still gates).
• CHANGELOG Unreleased entry.

Tests (23 new; suite 306 -> 329 passed, 1 skipped)

The three contract behaviors are asserted directly:

• no-regression exits 0 with a MATCH summary (test_no_regression_exits_zero_with_match_summary)
• injected regression exits nonzero and names the drifted claim (test_injected_regression_exits_nonzero_and_names_the_drifted_claim, asserting the drifted claim id appears in both stderr and the Markdown)
• the Markdown summary is deterministic (test_markdown_summary_is_deterministic_across_runs and a unit-level determinism test)

Plus: dropped-claim regression, drift->unverifiable regression, improvement/new not failing, tampered-baseline rejection, wrong-kind rejection, order-independent seal, JSON output, --out file behavior, and cells_from_latest skipping unknown statuses rather than guessing.

Test evidence

$ python -m pytest -q
329 passed, 1 skipped in 8.46s

$ python -m mypy src/crucible
Success: no issues found in 42 source files

$ python -m ruff check src/crucible/ci_gate.py src/crucible/ci_report.py src/crucible/ci_cmd.py tests/test_ci_gate.py tests/test_cli_ci.py
All checks passed!

Shipped vs specced

Fully shipped and tested; nothing deferred to a spec. Zero-dependency and Fair-Source posture preserved (stdlib only). Zero em-dashes / en-dashes. Experimental nothing; all behavior is exercised by tests.

Honest limitations

• A baseline is committed and refreshed by a maintainer when the posture legitimately advances; the gate does not auto-advance the baseline (by design: an auto-advancing baseline cannot catch a slow regression).
• The gate compares verdict status (MATCH/DRIFT/UNVERIFIABLE), not margin deltas within a status; a claim degrading while still MATCH reads as held. crucible drift remains the margin-delta view.
• The gate keys on (thesis_id, claim_id); a claim whose id changes (e.g. its content hash shifts) is seen as a drop + a new claim, which is the fail-closed reading.
• Pre-existing, out-of-scope: tests/test_browser_evidence.py has an unrelated ruff I001 import-order warning that exists on main; not touched here.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com

🤖 Generated with Claude Code