v1.2.0 - Headless L1 runner

Adds l1-run, a headless L1 runner that drives the full L1 loop (author,
adversarial judge, apply) through direct single-shot LLM API calls, so L1
ingest can run unattended from cron or a CI step. Additive: the existing
runners (Claude Code, Codex CLI, Copilot projection), the canonical contracts,
the adversarial gate and the state machine are unchanged.

Part of the vscode_abap_remote_fs integration
(issue #473):
this is the programmatic inference path a proxy or an editor command can drive.

• pipeline.py l1-run: batch loop via direct LLM calls, no chat runner and no
  VS Code required; exit codes suited to cron and CI.
• llm_client: two profiles (author and judge), both anthropic (Messages)
  and openai (chat completions) wire shapes, api keys read from the
  environment only, standard-library urllib with no new dependencies.
• core/docs/15-headless-l1-runner.md: usage, configuration and guarantees.

See the CHANGELOG for the full list.

v1.1.0 - ABAP FS integration groundwork

v1.1.0 - ABAP FS integration groundwork

Adds the abap_wiki side of the vscode_abap_remote_fs integration
(issue marcellourbani/vscode_abap_remote_fs#473):

• pipeline.py l0-run: one-shot deterministic L0 bootstrap
• Copilot projection: .github/agents/*.agent.md, generated from the
  same canonical contracts as .claude/agents/ and .agents/agents/
• core/docs/14-abap-fs-integration.md: end-to-end integration guide

v1.0.0 - First public release

First public release.

Added

• Engine (L0/L1): deterministic L0 inventory from a TADIR export (init-db, import-tadir, resolve-sources, ingest-l0, enqueue-l1, ingest-metadata); gated L1 code analysis with an independent adversarial judge, fail-closed promotion (no --force), idempotent apply, per-batch commits, deterministic include-edge derivation, and full loop recovery (claim, submit-author, submit-verdict, apply, recover, project, reopen-l1, rerender-pages, link-includes, requeue-skipped, retry-reset, gc-runs).
• L2 functional process: slice manifests with a mandatory real owner, gap discovery and multi-source auto-research, expert questionnaires with pre-filled hypotheses, answer capture, functional synthesis with line- and evidence-anchored claims, an independent fidelity gate, and gated promotion.
• State and views: SQLite as the single source of truth; wiki pages, indexes, dashboard, membership, Excel export and the operation log are regenerated projections.
• Quality guardrails: encoding check, mandatory three-part context headers on every engine code file, environment doctor with a fail-closed staged secret scan in the pre-commit hook, agent-contract synchronization check, wiki lint, and a 450+ case unit suite.
• Agent contracts and skills for Claude Code and Codex CLI: author, adversarial judge, functional researcher/author/gate agents; ingest, query, lint, slice and answer-capture skills.
• Demos and evidence: a one-command, zero-token L0 demo on a bundled synthetic package; a token-saving example knowledge base with measured compression; and a committed model-comparison benchmark across Claude author/judge pairings.
• Documentation: architecture, pipeline, adversarial gate, L2 process, lessons learned, runbook, testing, autonomous loop, section semantics, first-clone SAP input guide, roadmap, agent runtime and cost, FAQ.
• Community files: README, CONTRIBUTING, SECURITY, CODE_OF_CONDUCT, issue and PR templates, dependabot, CI.