CodeClone 2.1.0a1: intent-first structural change control, persistent engineering context, agent workflow evidence, platform self-observability, and broader IDE/agent integration.

CodeClone 2.1 introduces intent-first structural change control, persistent engineering context, agent workflow
evidence, platform self-observability, and broader IDE/agent integration.

Added

• Structural Change Controller with start_controlled_change / finish_controlled_change, bounded edit scope,
  blast-radius checks, patch verification, claim validation, multi-agent intent coordination, and deterministic review
  receipts.
• Live Implementation Context via get_implementation_context, including bounded structural context, call
  relationships, contract-oriented truth maps, freshness, test anchors, and active intent boundaries. Context remains
  read-only and never authorizes edits.
• Engineering Memory, Trajectory Memory, Patch Trail, and Experience Layer for typed repository
  knowledge, historical agent workflows, change evidence, reusable patterns, and human-governed promotion.
• Semantic retrieval with optional LanceDB hybrid search, FTS5/BM25, vector search, and deterministic Reciprocal
  Rank Fusion.
• Platform Observability for development-time tracing of CLI, MCP, analysis phases, database activity, semantic
  indexing, worker chains, memory/CPU use, MCP payload pressure, and costly no-ops.
• Corpus Analytics for offline intent clustering, interpretability, versioned profiles, sweep comparison, maintainer
  selection, and inspectable JSON/HTML outputs.
• Module Map as a deterministic report-only package/module graph with cycle, hub, overloaded-module, and
  unwind-candidate views.
• Guided Finding Review as a prioritized report-only review queue with shared finding cards, filters, progress
  tracking, and reviewed-state persistence.
• Native agent and IDE integrations for VS Code, Claude Desktop, Claude Code, Codex, and Cursor, including
  governance, audit, memory, trajectory, and structural-review workflows.
• codeclone setup — lazy-loaded CLI readiness surface (status, doctor, plan, apply, wizard) with
  capability-aware snapshots, read-only diff preview, and bounded pyproject.toml / .gitignore merges (no MCP
  intent, no baseline or report writes). The CLI also gains a guided --help tour.
• Expanded controller, memory, trajectory, analytics, semantic-search, observability, blast-radius, patch-verification,
  and diagnostic CLI/MCP surfaces. The default MCP server surface is now 38 tools (40 when VS Code enables the
  IDE governance channel).
• Reorganized documentation into a task-oriented site (concepts, guides, reference, integrations) with unified
  integration guidance and explicit edition tiers.
• MCP schemas now include parameter descriptions, deterministic next_tool guidance, token-budget tracking, workspace
  hygiene warnings, and documentation-contract linting.
• MCP response governance now advertises context_governance metadata for bounded agent replies. Workflow, memory, and
  implementation-context responses preserve mandatory control facts inline, compact recoverable evidence under
  partial_enforce, disclose omitted lanes, and expose exact drill-down through durable receipt, Patch Trail, blast
  artifact, memory continuation, and implementation-context page retrieval.

Contract changes

• Cache schema advanced to 2.9 for the rebuildable per-function relationship-fact projection and to 2.10 for
  intra-module, class-method, and receiver-aware call resolution.
• Engineering Memory schema advanced to 1.7 for trajectory and Patch Trail evidence.
• Semantic index format advanced to 3 for LanceDB rows with source_revision; existing semantic sidecars should be
  rebuilt.
• Corpus Analytics store schema advanced to 1.2.
• Corpus Analytics JSON export schema advanced through 1.2 and 1.3.
• Corpus Analytics representation contract advanced to 3.
• Corpus Analytics control-plane contract introduced at 1.0.
• MCP response governance contract introduced at 1.0 with deterministic utf8_bytes_div_4_v1 context-unit
  estimation and explicit observe / partial_enforce modes.
• compare_runs parameters are unified to before_run_id / after_run_id.
• derived.module_map and derived.review_queue remain report-only projections excluded from the integrity digest;
  they add no analysis pass, metrics family, or report schema bump.
• Live Implementation Context relationship facts remain off the canonical report and do not change canonical report
  identity.

Changed

• Default project workspace moved from .cache/codeclone/ to .codeclone/; legacy paths emit a migration warning.
• Documentation builds now use Zensical with strict clean builds.
• pydantic is now a base dependency.
• LCOM4 excludes Protocol methods and Pydantic validation/serialization hooks; computed_field remains included.
• Repository coverage is enforced at >=99%.

Performance

• Relationship facts are collected during the primary module walk instead of a second traversal (cold-cache
  phase_relationship −28%), and reachability facts replay captured handler nodes instead of a third full AST pass.
• MCP hot paths are bounded: findings are paged before decoration, hotspot and intent-scoped blast projections are
  capped, analysis no longer round-trips through the report, and cache loads retain less transient memory.

Fixed

• Engineering Memory writes are durable, batch ingestion is atomic, and memory/trajectory/Patch Trail lifecycle
  handling avoids premature staleness, duplicate projections, stale workflow rows, and broken evidence links.
• Best-effort audit and memory-proposal failures are now observable instead of silently swallowed.
• Implementation-context misses return a compact actionable payload instead of empty scaffolding.
• Workspace hygiene, intent attribution, continuation of owned work, queue handling, and recoverable-intent behavior
  were corrected.
• Patch verification now rejects identical before/after runs where required, surfaces health regressions, and warns on
  overstated review claims.
• Semantic retrieval now preserves lexical/vector relevance, avoids source crowding, loads embeddings lazily, and
  coalesces redundant projection work.
• Blast-radius graph logic moved into codeclone/analysis/blast_radius.py, removing the CLI-to-MCP dependency
  violation.
• respect_pyproject=false no longer reports golden-fixture clone groups as false new regressions.
• Documentation URLs, integration references, and contract tests were aligned with the reorganized site.