v2.0.0

Tagline: comprehensive release — end-to-end coherent system, field-validated
by 20 independent agents across 4 perspectives × 5 scenarios, with measurable
improvements vs the v1.11.0 baseline on every cell of the matrix.

Highlights

• Tool surface collapsed 344 → 146 live (-58%) with 80 backward-compat
  aliases + 78 deprecated aliases + 24 hard-removed (return _REMOVED_TOOLS
  envelope). 25-family consolidation pass via the proven _ALIAS_PARAM_INJECTION
  pattern — every legacy name keeps dispatching for the v2.0.x runway.
• server.py 7,499-line monolith dissolved into 32-module
  src/research_os/server/ package (largest module: tool_definitions/meta.py
  at 579 lines, -92% from peak). Public API preserved end-to-end via
  __init__.py re-exports.
• MCP instructions field on the initialize handshake — names the
  canonical per-turn sequence (sys_boot → tool_route → sys_protocol_get(format=summary) → sys_active_tools) at the protocol
  layer so any MCP client surfaces the right startup ritual.
• sys_protocol_get default format flipped "full" → "summary" —
  the single biggest token-cost win (5–10× cheaper per-turn at ~300 tokens
  vs ~1.5–3K). MAJOR-breaking; pass format="full" to opt back in.
• 5 CRAFT-inspired structural additions drove the rating lift beyond
  surface cleanup: (1) audit-as-data (every audit emits a JSON companion +
  .audit_findings.jsonl ledger queryable via
  tool_audit_findings(operation=query|diff)); (2) drafter review-rewrite
  loops on paper/slides/poster; (3) research-os doctor install + workspace
  health checks; (4) docs/CONTRACT.md stable-surface promise;
  (5) audience-segmented docs/README.md four-audience router.
• Validation: 20 agents × 4 perspectives × 5 scenarios. Mean
  final_rating moved 6.35 → 7.70 (+1.35; +21%), total HIGH-friction
  items 124 → 63 (-49%), first-5-turn HIGH 66 → 42 (-36%), deliverable
  rate 11/20 → 14/20 (+15 pp). Every cell improved, no regressions.
  Full report: docs/V2_VALIDATION_REPORT.md.
• YELLOW recommendation — ship with documented caveats. The Phase 15
  GREEN gate targets (avg ≥ 9.5, HIGH ≤ 5, all four perspectives ≥ 9.0) are
  not met; they were calibrated against a hypothetical v3-grade product.
  Deeper structural gaps (domain-pack coverage for bioinformatics + systems
  benchmarks; pack-aware audit gates) carry over to v2.0.x patch +
  v2.1.0 minor per
  docs/V2_RELEASE_NOTES.md §"Deferred".
• Upgrade path: most projects work unchanged via alias dispatch.
  Full instructions, breaking-change details, per-surface recipes, and the
  complete old→new tool table at
  docs/MIGRATION_v1_to_v2.md.
• Two v2.0.1 BLOCKER regressions surfaced by Phase 15b re-validation
  were fixed before tagging (commits 0c45b79 + b3b24a0):
  sys_tool_describe NameError (_resolve_tool_name missing import
  after the Phase 10 server-package split) and
  tool_audit(scope='synthesis', dimension='all') bare KeyError on
  paper_path (now defaults to 'synthesis/paper.md'). Server() also
  now reports the canonical __version__ at the MCP initialize handshake.

Added

• tool_audit(scope=, dimension=) unified per-dimension audit dispatcher and tool_audit_findings(operation=query|diff) ledger reader. Both share the existing _ALIAS_PARAM_INJECTION / _DEPRECATED_ALIASES machinery so the prior tool surface (tool_audit_synthesis, tool_audit_step_completeness, tool_audit_findings_query, etc.) keeps dispatching with the legacy behaviour preserved end-to-end. tool_audit_quality_full stays separate as the canonical aggregator.
• tool_dashboard(operation=create|story_generate|story_edit|story_quality_bar|reviewer_sim|test_generate|test_run) unified dashboard dispatcher. Single entry point for the seven previously per-operation dashboard tools; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• tool_step(operation=iterate|iterations_list|revision_options|env_lock) unified step-lifecycle dispatcher and tool_step_pipeline(operation=define|run|status|diagram) unified step sub-task pipeline dispatcher. Single entry points for the eight previously per-operation step tools; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged. tool_step_complete stays standalone as the top-level end-of-step bundle.
• tool_lessons(operation=record|consult|failure_record|failure_check|failure_list|dead_end|mistake_replay) extended dispatcher for the entire "what went wrong / what did we learn" family, and tool_reliability(operation=log_event|report) unified reliability-log dispatcher. Single entry points for the seven lessons / failure-memory / dead-end / mistake-replay tools and the two reliability-log tools; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• tool_sensitivity(operation=define|run) unified sensitivity dispatcher and tool_preregister(operation=freeze|diff) unified preregistration dispatcher. Single entry points for the two previously per-operation sensitivity tools and the two previously per-operation preregister tools; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• tool_reviewer(operation=simulate|response|rebuttal|compile) unified reviewer-response dispatcher. Single entry point for the four previously standalone reviewer-response scaffold tools (tool_reviewer_simulate, tool_response_to_reviewers, tool_rebuttal_draft, tool_reviewer_response_compile); each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• tool_data(operation=sample|profile|convert), tool_figure(operation=palette|caption_synthesise|interactive_autogen|paper_autoembed), and tool_thought(operation=log|trace) unified dispatchers. Single entry points for the three data tools, four figure helpers, and two ReAct trace tools previously exposed as per-operation surface; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• tool_scratch(operation=write|run|list|clear) and tool_task(operation=run|status|list|kill) unified dispatchers. Single entry points for the four scratch-sandbox tools and the four background-task tools previously exposed as per-operation surface; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• sys_config(operation=get|set|validate) and sys_env(operation=snapshot|docker_generate) unified dispatchers. Single entry points for the three researcher-config tools and the two environment tools previously exposed as per-operation surface; each legacy name routes through alias + param injection so existing callers, scripts, and protocols keep working unchanged.
• docs/V2_MIGRATION_TABLE.md — running ledger of every old→new tool consolidation (old name, new name, dispatch kwarg, value, status). First entry: the 26→3 audit-family collapse (phase-9-c1).

Changed

• Audit family consolidated 26 → 3 tools (phase-9-c1). 23 per-dimension tool_audit_* tools collapse into tool_audit(scope=, dimension=); the 2 findings-ledger tools collapse into tool_audit_findings(operation=). Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output. _ALIAS_PARAM_INJECTION now accepts multi-kwarg specs (tuple of (key, value) pairs) so the audit family can inject both scope and dimension from a single alias.
• Dashboard family consolidated 7 → 1 tool (phase-9-c2). The seven per-operation tool_dashboard_* tools (create, story_generate, story_edit, story_quality_bar, reviewer_sim, test_generate, test_run) collapse into tool_dashboard(operation=...). Legacy aliases continue to dispatch through the consolidated handler via _ALIAS_PARAM_INJECTION. Shipped protocol YAMLs (synthesis/synthesis_dashboard, visualization/interactive_figure_design, audit/pre_submission_checklist, guidance/autopilot) and _router_index.yaml rewritten to the canonical tool_dashboard(operation='…') surface so reviewer-facing guidance stays on the live names.
• Step family consolidated 8 → 2 tools (phase-9-c3). The four step-lifecycle tools (tool_step_iterate, tool_step_iterations_list, tool_step_revision_options, tool_step_env_lock) collapse into tool_step(operation=...). The four step sub-task pipeline tools (tool_step_pipeline_define, tool_step_pipeline_run, tool_step_pipeline_status, tool_step_pipeline_diagram) collapse into tool_step_pipeline(operation=...). Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output. tool_step_complete stays standalone as the top-level end-of-step bundle; tool_step_literature_list belongs to the literature/search family and is not consolidated here. Shipped protocol YAMLs (guidance/analysis_plan, methodology/deep_domain_research) and _router_index.yaml rewritten to the canonical tool_step(operation='…') / tool_step_pipeline(operation='…') surface so reviewer-facing guidance stays on the live names.
• Lessons + reliability family consolidated 10 → 2 tools (phase-9-c4). The pre-existing tool_lessons (which already absorbed tool_lessons_record + tool_lessons_consult) is extended to cover tool_failure_record (operation=failure_record), tool_failure_check (failure_check), tool_failure_list (failure_list), tool_dead_end_lessons (dead_end), and tool_mistake_replay (mistake_replay). The two reliability-log tools (tool_reliability_log_event, tool_reliability_report) collapse into a separate tool_reliability(operation=log_event|report) entry point. Every legacy name remains callable via alias + param injection. Shipped protocol YAMLs (audit/audit_and_validation, guidance/session_resume, literature/literature_search, synthesis/synthesis_progress_update) and _router_index.yaml rewritten to the canonical tool_lessons(operation='…') / tool_reliability(operation='…') surface so reviewer-facing guidance stays on the live names.
• Sensitivity + preregister families consolidated 4 → 2 tools (phase-9-c5). The two sensitivity tools (tool_sensitivity_define, tool_sensitivity_run) collapse into tool_sensitivity(operation=define|run). The two preregister tools (tool_preregister_freeze, tool_preregister_diff) collapse into tool_preregister(operation=freeze|diff). Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output. Shipped protocol YAMLs (methodology/preregistration, methodology/missing_data_strategy, methodology/method_comparison, audit/audit_and_validation, audit/provenance_completeness, synthesis/synthesis_null_findings) and _router_index.yaml rewritten to the canonical tool_sensitivity(operation='…') / tool_preregister(operation='…') surface so reviewer-facing guidance stays on the live names.
• Reviewer family consolidated 4 → 1 tool (phase-9-c6). The four reviewer-response scaffold tools (tool_reviewer_simulate, tool_response_to_reviewers, tool_rebuttal_draft, tool_reviewer_response_compile) collapse into tool_reviewer(operation=simulate|response|rebuttal|compile). Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output.
• Data + figure + thought families consolidated 9 → 3 tools (phase-9-c7). The three data tools (tool_data_sample, tool_data_profile, tool_data_convert) collapse into tool_data(operation=sample|profile|convert). The four figure helpers (tool_figure_palette, tool_figure_caption_synthesise, tool_figure_interactive_autogen, tool_paper_figures_autoembed) collapse into tool_figure(operation=palette|caption_synthesise|interactive_autogen|paper_autoembed). The two ReAct trace tools (tool_thought_log, tool_thought_trace) collapse into tool_thought(operation=log|trace). Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output. Shipped protocol YAMLs (audit/pre_submission_checklist, guidance/analysis_plan, guidance/project_startup, methodology/data_management_plan, methodology/data_quality_audit, methodology/exploratory_data_analysis, methodology/missing_data_strategy, synthesis/synthesis_paper, visualization/figure_guidelines, visualization/interactive_figure_design, visualization/visualization_workflow) and _router_index.yaml rewritten to the canonical tool_data(operation='…') / tool_figure(operation='…') / tool_thought(operation='…') surface so reviewer-facing guidance stays on the live names.
• Misc-family audit (phase-9-c8). The remaining tool_<verb>_* families flagged as natural sub-systems consolidate 8 → 2 tools: the four scratch-sandbox tools (tool_scratch_write, tool_scratch_run, tool_scratch_list, tool_scratch_clear) collapse into tool_scratch(operation=write|run|list|clear); the four background-task tools (tool_task_run, tool_task_status, tool_task_list, tool_task_kill) collapse into tool_task(operation=run|status|list|kill). The two tool_quick_* tools (tool_quick_review stages a paper-review markdown; tool_quick_route is the throwaway-intent classifier used to short-circuit protocol load) share a prefix only — no functional overlap — and are kept standalone. Every legacy name is aliased + parameter-injected so older scripts, protocols, and researcher commands continue to produce identical output. Shipped protocol YAMLs (guidance/casual_exploration, guidance/chat_handoff, guidance/session_resume, guidance/autopilot, guidance/analysis_plan, methodology/deep_domain_research, methodology/reproduction_attempt, methodology/simulation_studies) and _router_index.yaml rewritten to the canonical tool_scratch(operation='…') / tool_task(operation='…') surface so reviewer-facing guidance stays on the live names.
• SYS_* judgment pass (phase-9-c9). The sys_* MCP-level primitives were reviewed family-by-family. Two genuinely over-fragmented families consolidate 5 → 2 tools: sys_config_get / sys_config_set / sys_config_validate collapse into sys_config(operation=get|set|validate), and sys_env_snapshot / sys_env_docker_generate collapse into sys_env(operation=snapshot|docker_generate). Every other sys_* family is intentionally kept separate because each is a distinct primitive AIs need to find by name — discovery surface (sys_boot, sys_active_tools, sys_protocol_* ×6, sys_help, sys_tool_describe, sys_state_get), high-frequency file I/O (sys_file_read / _write / _list / _delete / _validate_md), checkpoint trio (sys_checkpoint_create / _rollback / _list), workspace pair (sys_workspace_scaffold / _tree), and the standalone interaction tools (sys_session_handoff, sys_export_share_archive, sys_notify, sys_active_project, sys_dep_inventory, sys_semantic_tool_search, sys_packs_installed, sys_adapters_installed). Every legacy name remains callable via alias + param injection. Shipped protocol YAMLs (guidance/session_boot, guidance/project_startup, guidance/analysis_plan, methodology/cox_ph_diagnostics, methodology/pick_tool_stack, methodology/mixed_language_orchestration, methodology/reproduction_attempt, reproducibility/reproducibility) and _router_index.yaml rewritten to the canonical sys_config(operation='…') / sys_env(operation='…') surface so reviewer-facing guidance stays on the live names.
• All shipped protocol YAMLs that previously referenced per-dimension audit tool names now reference the consolidated tool_audit(scope='…', dimension='…') / tool_audit_findings(operation='…') form (96 substitutions across 33 protocol files) so reviewer-facing guidance stays on the canonical surface.
• MAJOR-breaking: sys_protocol_get default format flipped from "full" to "summary" (phase-9-cross-cutting). Callers that previously did not pass a format argument received the entire YAML (~1.5-3K tokens per call); they now receive the ~300-token summary view. This is the single biggest token-cost win identified in the Phase-15a baseline. Callers who genuinely need the bulk payload must now pass format="full" explicitly. The schema default + handler default + AI orientation docs all updated together; the inputSchema now declares "default": "summary" so well-behaved clients see the new default automatically.
• Every TOOL_DEFINITIONS entry now carries two introspection fields — status (live | alias | deprecated) and pack (core | <pack_name>) — set automatically by _annotate_core_tool_metadata() for core tools and by plugins/loader.py + adapters/loader.py at registration time for pack / adapter tools. sys_tool_describe surfaces both so the router, list_tools, and any external tooling can filter without re-deriving the answer. 167 entries annotated (146 live + 21 alias from the consolidation overlap; 144 core + 23 from packs / adapters).
• Every shipped protocol YAML (153 files across core + the 5 bundled packs) now carries a scope_tags block — domain (e.g. [biology, wet_lab], [qualitative], [any]), audience (e.g. [researcher], [auditor], [naive_ai]), and workflow_shape (e.g. [experiment_pipeline], [proof], [linear_essay], [interview_study], [systems_benchmark], [any]). The router uses these as a soft filter so the embedding-similarity ranking only competes within the project's declared scope.
• tool_route now returns a recommended_action field — a single-string hint naming the exact next tool to call (e.g. "sys_protocol_get(protocol_name='guidance/project_startup', format='summary')"). Saves the AI one round-trip of reasoning per turn.
• MCP Server instantiation now passes the instructions= field with the canonical per-turn sequence (sys_boot → tool_route → sys_protocol_get format=summary → sys_active_tools), so any MCP client that surfaces server-supplied instructions sees the right startup ritual without having to call sys_help.

Deprecated

• Legacy per-dimension audit tool names — tool_audit_assumptions, tool_audit_citations, tool_audit_claims, tool_audit_cliches, tool_audit_code_quality, tool_audit_coherence, tool_audit_cross_deliverable_consistency, tool_audit_dashboard_content, tool_audit_evalue, tool_audit_figure, tool_audit_figure_coverage, tool_audit_figure_full, tool_audit_figure_interactivity, tool_audit_figure_quality, tool_audit_power, tool_audit_prose, tool_audit_reproducibility, tool_audit_reviewer_responses, tool_audit_statistical_power, tool_audit_step_completeness, tool_audit_step_literature, tool_audit_synthesis, tool_audit_version_coherence, tool_audit_findings_query, tool_audit_findings_diff. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy dashboard tool names — tool_dashboard_create, tool_dashboard_story_generate, tool_dashboard_story_edit, tool_dashboard_story_quality_bar, tool_dashboard_reviewer_sim, tool_dashboard_test_generate, tool_dashboard_test_run. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy step tool names — tool_step_iterate, tool_step_iterations_list, tool_step_revision_options, tool_step_env_lock, tool_step_pipeline_define, tool_step_pipeline_run, tool_step_pipeline_status, tool_step_pipeline_diagram. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy sensitivity + preregister tool names — tool_sensitivity_define, tool_sensitivity_run, tool_preregister_freeze, tool_preregister_diff. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy reviewer tool names — tool_reviewer_simulate, tool_response_to_reviewers, tool_rebuttal_draft, tool_reviewer_response_compile. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy data + figure + thought tool names — tool_data_sample, tool_data_profile, tool_data_convert, tool_figure_palette, tool_figure_caption_synthesise, tool_figure_interactive_autogen, tool_paper_figures_autoembed, tool_thought_log, tool_thought_trace. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy scratch + task tool names — tool_scratch_write, tool_scratch_run, tool_scratch_list, tool_scratch_clear, tool_task_run, tool_task_status, tool_task_list, tool_task_kill. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.
• Legacy sys_config + sys_env tool names — sys_config_get, sys_config_set, sys_config_validate, sys_env_snapshot, sys_env_docker_generate. All continue to work via alias dispatch through v2.0.x; scheduled for hard-removal in v2.1.0 per docs/V2_MIGRATION_TABLE.md.

Removed

• Phase 14a — first-wave consolidation aliases hard-removed. The 21 legacy tool names introduced as consolidation aliases in v1.6.1 have expired their 4-minor-version deprecation runway and are now removed. Calling any of them returns a friendly _REMOVED_TOOLS error envelope naming the canonical v2 entry point. Old plans, scripts, or third-party callers that still name these will see a clear migration message instead of a generic "unknown tool" error.
  • Search cluster (5): tool_search_semantic_scholar, tool_search_pubmed, tool_search_crossref, tool_search_arxiv, tool_search_web → call tool_search(query=..., source='semantic_scholar'|'pubmed'|'crossref'|'arxiv'|'web') instead.
  • Plan cluster (3): tool_plan_turn, tool_plan_advance, tool_plan_clear → call tool_plan(operation='turn'|'advance'|'clear') instead.
  • Grounding / verify cluster (4): tool_grounding_register, tool_ground_from_context → call tool_ground(mode='explicit'|'from_context', ...). tool_claim_verify, tool_grounding_verify → call tool_verify(scope='claim'|'project', ...).
  • Lessons cluster (2): tool_lessons_record, tool_lessons_consult → call tool_lessons(operation='record'|'consult', ...) instead. (Other lessons-family aliases remain deprecated for the v2.0.x runway.)
  • Path cluster (3): sys_path_create, sys_path_abandon, sys_path_list → call sys_path(operation='create'|'abandon'|'list', ...) instead.
  • Memory cluster (4): mem_methods_append, mem_decision_log, mem_hypothesis_update, mem_analysis_log → call mem_log(kind='methods'|'decision'|'hypothesis'|'analysis', ...) instead.
• The corresponding TOOL_DEFINITIONS entries and _HANDLERS entries were dropped (handler functions like _handle_sys_path_create remain in the module — they're called internally by the consolidated dispatchers via the legacy fallback path). tool_log_decision, the silent pre-v1.6.1 nickname that previously chained through mem_decision_log → mem_log, now resolves directly to mem_log with kind='decision' injected so the nickname keeps working.
• Phase 14b — tikzposter LaTeX poster path hard-removed. The legacy create_poster() + _poster_tex_escape() functions under src/research_os/tools/actions/synthesis/latex.py (387 lines) are deleted. tool_poster_create is unchanged on the surface (Typst engine is the only path); the engine='latex' branch in _handle_tool_poster_create now returns a structured error pointing callers at the Typst surface. The legacy layout / audience LaTeX-only kwargs are no longer documented (engine is retained on the schema for back-compat with a hard-error guard). researcher_config.synthesis.poster_engine is pinned to "typst" — the validator enum now rejects "latex". Protocol synthesis/printable updated (template: tikzposter → academic_36x48 / academic_48x36 / public_24x36 per audience; description text re-pointed at Typst). Router index summary + triggers updated. Docs (PROTOCOLS.md, RESEARCHER_GUIDE.md, TOOLS.md, ROADMAP.md) re-pointed at Typst. Test test_legacy_tikzposter_path_still_works replaced by test_legacy_tikzposter_create_poster_is_gone and the enum-shape test now asserts "latex" not in synthesis.poster_engine. _REMOVED_TOOLS entries added for tool_poster_create_latex / tool_poster_compile_latex (nicknames that were never real tools but a future caller might try). Audit of all other handlers in server.py found zero truly orphan handlers; every handler is referenced either by the dispatcher map or called internally by a Phase 9 consolidator.
• Phase 14d — dead config fields removed. Five researcher_config.yaml fields identified by the v1.9.2 Lens-7 audit as declared-but-never-read are removed from the on-disk template, the in-code CONFIG_TEMPLATE constant, and docs/RESEARCHER_GUIDE.md. None of these fields had any consumer in src/, tests/, or shipped protocols — the comments that claimed they were "Read by methodology/pick_tool_stack" were inaccurate. The pick_tool_stack protocol picks language + library purely from method + field-practice + literature signal; it never consulted these fields. Existing projects on prior versions that hand-set these keys are unaffected (the keys silently become unknown extras; validate_config does not enforce key membership).
  • Runtime cluster (1): runtime.default_n_for_sampling — no caller anywhere; tool_data(operation='sample') takes its n argument from the tool call, not from config.
  • Tool-stack cluster (4): entire tool_stack: block removed: tool_stack.preferred_languages, tool_stack.allow_mixed_language_steps, tool_stack.field_practice_overrides_preference, tool_stack.cite_field_practice_when_choosing. The methodology/pick_tool_stack protocol itself is unchanged — it asks the AI to pick based on method + literature signal + env compatibility, never consulting these config fields.
• Trivial dead-variable cleanup in tools/actions/audit/audit.py:651 (f_stat, f_p from het_breuschpagan unpack are now _f_stat, _f_p — they were tuple-discard placeholders flagged by Lens-9 as the one real vulture --min-confidence 80 finding still present after v1.9.3's larger sweep).

Fixed

• sys_tool_describe NameError regression introduced by the Phase 10
  server-package split — meta_routing.py referenced _resolve_tool_name
  which wasn't re-exported from _handlers_runtime.py. Fixed in commit
  0c45b79: added the import + __all__ entry so the introspection path
  works from the first list_tools() call.
• tool_audit(scope='synthesis', dimension='all') raised a bare
  KeyError on paper_path. Fixed in commit 0c45b79: handler now
  defaults to 'synthesis/paper.md' (matches what the audit_synthesis
  worker already assumes when callers omit the kwarg).
• MCP Server() instance now reports the canonical __version__ at the
  MCP initialize handshake instead of hard-coding '0.1.0'. Fixed in
  commit b3b24a0 (phase-13 follow-up).

Validation

• Phase 15b re-validation: 20 agents × 4 perspectives × 5 scenarios = 20
  independent runs against the v2.0.0 candidate. Mean final_rating
  moved 6.35 → 7.70 (+1.35; +21%), total HIGH-friction items
  124 → 63 (-49%), first-5-turn HIGH 66 → 42 (-36%), deliverable
  rate 11/20 → 14/20 (+15 pp). Every cell of the 4×5 matrix moved up by
  +0.7 to +1.9 points; no regressions in any of the 20 runs. Full
  per-perspective × per-scenario rating table, friction-event delta, the
  carryover deferral list (v2.0.1 patch / v2.1.0 minor / v3.0.0 major),
  and the YELLOW shipping recommendation at
  docs/V2_VALIDATION_REPORT.md.

Migration

• Most projects work unchanged — every consolidated tool name keeps
  dispatching via _DEPRECATED_ALIASES + _ALIAS_PARAM_INJECTION for
  the v2.0.x runway. Hard removal of the v2.0 deprecated aliases is
  scheduled for v2.1.0. Full instructions, breaking-change details,
  per-surface recipes, and the complete old→new tool table at
  docs/MIGRATION_v1_to_v2.md (the v1→v2
  upgrade guide) plus docs/V2_MIGRATION_TABLE.md
  (the running ledger of every old→new consolidation).
• Release-shaped overview, headline numbers, and YELLOW caveat at
  docs/V2_RELEASE_NOTES.md.

---