…review feedback

- Added MongoDB implementation for database interfaces
- Implemented Advanced Search (Vector + Hybrid)
- Added JSON-LD context, enrichment, and integrity signing
- Fixed initialization issue in server setup
- Corrected analytics aggregation logic
- Improved database connection resilience
- Refactored shared logic for size analysis

All 5 base challenges + 4 wow factors:
- BASE 1: SIPREC adapter (folder drop ingestion)
- BASE 2: MQTT/UNS bridge (real-time event streaming)
- BASE 3: Real-time analytics dashboard + ingestion UI
- BASE 4: Teams adapter (MS Graph callRecord)
- BASE 5: Neo4j consumer (relational graph mapping)
- WOW 1: JSON-LD-ex semantic enrichment (confidence algebra, integrity, provenance)
- WOW 2: WhatsApp adapter (chat export parser)
- WOW 3: RAG intelligence (Whisper GPU + LLaMA/Groq)
- WOW 4: Real-time dashboard with Neo4j graph, RAG chat, JSON-LD inspector

Infrastructure: Docker (MongoDB, Neo4j, Mosquitto, ChromaDB), Python sidecars
All components implemented as plugins - zero core modifications

Spec compliance (draft-ietf-vcon-vcon-core-02 / v0.4.0):
- Rename must_support → critical, appended → amended across all code
- Update VConVersion to '0.4.0', add SessionId type {local, remote}
- Fix Analysis vendor as required (not optional) throughout
- Add 9 missing IVConQueries interface methods (tags, updateVCon, searchCount)
- Fix mongo-queries.ts and queries.ts to use new field names
- Fix validation.ts version check and validateMustSupport → validateCritical

Import scripts:
- Add scripts/import-vcon-files.ts — bulk importer for real .vcon files
  (handles v0.0.1 Strolid format, concurrency pool, skip-existing, error log)
- Add scripts/import-demo-conversations.ts — demo data importer
- Add DEMO.md — Claude Desktop demo guide with copy-paste prompts

Documentation (doc-review fixes):
- README: badge and feature bullet updated to draft-02/v0.4.0
- README: Claude Desktop config now shows SUPABASE_SERVICE_ROLE_KEY
- docs/reference/vcon-spec.md: all v0.3.0/must_support/appended/mimetype refs updated
- docs/reference/CORRECTED_SCHEMA.md: DDL fixed (amended/critical), warning banner added
- docs/reference/IMPLEMENTATION_CORRECTIONS.md: spec ref updated to -02
- docs/guide/getting-started.md: dead links removed, version strings fixed
- scripts/README.md: import-vcon-files.ts documented
- CLAUDE.md: absolute spec path replaced with repo-relative + IETF datatracker URL

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…dout pollution

ConsoleSpanExporter and ConsoleMetricExporter write span/metric data to
stdout by default. In MCP stdio mode Claude Desktop reads stdout as
JSON-RPC, so OTEL output caused SyntaxError on every tool call.

Replace both with NullSpanExporter and NullMetricExporter (no-op
implementations) as the fallback when OTEL_EXPORTER_TYPE != 'otlp'.
Set OTEL_EXPORTER_TYPE=otlp with a real collector to get telemetry.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…lback

When the search_vcons_by_tags RPC is unavailable, the manual fallback
collects matching vcon_ids and then looks up their UUIDs with a single
.in('id', [...]) query. With 13k+ matching IDs (e.g. source:consig)
the generated URL exceeds the server's URI length limit.

Fix: split the .in() lookup into batches of 500 IDs to keep each
request's URL within limits.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Both search_vcons_semantic and search_vcons_hybrid were throwing
"Embedding generation not yet implemented" when given a plain text query.
Now they call OpenAI text-embedding-3-small (384 dimensions) via fetch
when no pre-computed embedding is provided.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ts_headline with HighlightAll=TRUE returns the entire document body —
up to 97KB per transcript. 50 results × ~40KB = well over the 1MB
MCP tool result size limit, causing all search_vcons_content calls to fail.

Two-layer fix:
1. New migration (20260414000000): replace HighlightAll=TRUE with
   MaxFragments=3, MaxWords=25 in search_vcons_keyword RPC so the DB
   returns compact context windows (~300-700 bytes each)
2. Handler safety cap: truncate any snippet over 500 chars in
   SearchVConsContentHandler to guard against future regressions

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

All three search RPCs (keyword, semantic, hybrid) were returning vcons.id
(the internal database PK) as vcon_id. The get_vcon tool looks up by
vcons.uuid (the external vCon spec identifier) — a different column.
Every search → get_vcon call failed with "vCon not found".

Fix per RPC:
- keyword: add internal_id to base CTE for the vcon_tags_mv join;
  JOIN vcons in party/dialog/analysis branches to get v.uuid
- semantic: JOIN vcons v ON v.id = e.vcon_id; return v.uuid
- hybrid: change final SELECT from v.id to v.uuid (joins already
  use v.id correctly via sem/kw CTEs)

Verified: returned vcon_id now matches vcons.uuid for all three RPCs.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

get_vcon now supports response_format:
- "full" (default): complete vCon with all analysis bodies
- "summary": metadata + parties + only summary-type analysis (~500 bytes
  vs 82KB+ for full transcripts) — use this when browsing search results
- "metadata": vCon metadata + parties only, no analysis/dialog

Updated tool descriptions to guide Claude:
- hybrid search: explain keyword_score=0 means semantic-only match (not
  that billing terms are absent), recommend get_vcon(response_format="summary")
- content search: recommend summary format for follow-up get_vcon calls

This prevents Claude from loading 82KB transcript JSON per vCon when
summarizing search results (10 full vCons = 820KB context overhead).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ints

Resolved conflicts between our v0.4.0 spec work and upstream changes:

Type relaxations (upstream PRs #35, #36):
- vcon.ts: VConVersion deprecated; vcon field is now optional string
- vcon.ts: disposition accepts any string (was DialogDisposition enum)
- vcon.ts: session_id is string for compatibility (v0.4.0 spec uses object)
- vcon.ts: Analysis.body is unknown (was string)
- vcon.ts: isValidDisposition returns boolean, accepts any non-empty string
- validation.ts: removed strict '0.4.0' version check

Parallelization (upstream PR #33):
- queries.ts: getVCon now uses Promise.all for parties/dialog/analysis/attachments
- queries.ts: createVCon inserts dialog/analysis/attachments in parallel
- queries.ts: createVCon deletes existing child rows before insert (upsert fix)
- database-inspector.ts: fixed indentation + kept includeCounts row count logic

Our changes preserved:
- queries.ts: SupabaseVConQueries implements IVConQueries (interface kept)
- queries.ts: v0.4.0 field names: critical (not must_support), '0.4.0' version
- resources/index.ts: IVConQueries interface + deserializeBody both imported
- services/vcon-service.ts: IVConQueries interface + pLimit both imported
- mongo-queries.ts: cast body to string for JSON.parse (body now unknown)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Add backward-compatible export aliases for renamed Supabase* DB classes
  (DatabaseSizeAnalyzer, DatabaseAnalytics, DatabaseInspector)
- Fix tag resource endpoint to use JSON.parse instead of deserializeBody,
  which left encoding='json' bodies unparsed and caused character iteration
- Update vcon-service tests: vcon version string 0.3.0 → 0.4.0
- Remove duplicate serial count query from getDatabaseShape that exhausted
  mock RPC calls before the relationships query
- Fix searchVCons integration test timeout: reduce 1500-item mock to 3
  items and spy on getVCon directly to avoid concurrent mock misalignment

All 629 tests now pass (5 skipped — intentional Mongo stubs).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>