[YoniRaviv]

Summary

Phase 5 wires the verifier/repairer (previously eval-only) into the real docpulse check / docpulse repair CLI, sourcing change intent from git, and turns verdicts + repairs into an honest CI exit code, a human report, and a dry-run companion-PR plan. Live git push / gh pr create are deferred to Phase 6.

New modules

• pipeline.py — CI-agnostic orchestrator: diff → suspects → verify → (repair) → RunResult, with the stale-index warning and per-suspect bundle seeding.
• context/ — ContextProvider protocol + GitContext (PR env vars → commit-message fallback → ticket IDs).
• destinations/ — Destination protocol + RepoMarkdownDestination: bottom-up section replacement, flag-tier comment, and the dry-run companion-PR FixPlan (branch/commit/PR-body + emitted gh/git commands), confidence-routed via Phase 4's route().
• report/summary.py — compute_exit_code + render_summary.
• command_runner.py — single shared injectable subprocess runner.

Changed

• diffing/git_diff.py — default three-dot base...head (merge-base) with --two-dot opt-in and auto-fallback when no common ancestor.
• cli.py — check now verifies + reports (exit 1 on drift); --suspects-only preserves the keyless suspect list; new repair command prints unified diffs + the dry-run PR plan, --write applies edits locally.

Locked decisions honored

• Two modes (check/repair); --suspects-only keyless path.
• Dry-run only — no push/PR this phase.
• Exit 1 iff a verdict is stale with confidence ≥ flag_threshold; unverified never fails; exit 2 = tool/setup error.
• Three-dot diff default.

Deferred items closed: RunResult.exit_code construction, stale-index warning, two-dot→three-dot, standalone repair --base.

Executed subagent-driven (per-task spec + quality review + a final integrated review). The review loop caught real bugs in plan-prescribed code: a splitlines() round-trip bug that silently deleted blank separators between doc sections (file-mutating core), a ticket-regex false-positive (UTF-8), a duplicated command-runner, and a render_summary "fixed" over-count (counted validation-failed repairs). All fixed.

Test Plan

• uv run pytest -q → 179 passed (from 137 baseline; +42 Phase 5 tests)
• uv run ruff check → clean
• Final integrated review: boundary rule holds (no gh/GitHub outside cli.py+destinations/), exit-code paths audited, replace_sections round-trips real multi-section docs, end-to-end section-id wiring consistent
• Live GitHub end-to-end run — Phase 6 manual gate (real push/PR not implemented here)

🤖 Generated with Claude Code