Mark production-ingestion-v1 epics as implemented in roadmap

Author: DiogoRibeiro7

ROADMAP.md

@@ -48,7 +48,15 @@ Issues #22 and #23 can land in parallel. #24 lands last and flips the gate on.
 
 [Milestone on GitHub](https://github.com/DiogoRibeiro7/llm-data-platform/milestone/2)
 
-**Why:** `llm_knowledge_ingestion` is explicitly an MVP today. Parser classes for PDF/HTML/Markdown raise `NotImplementedError`. Chunking is whitespace-window only with a hardcoded guard against any other strategy. The dedup module computes hashes but nothing compares them. Sources are local filesystem only. A single bad file aborts the whole run. This milestone closes those gaps.
+**Status (as of 2026-06-18): all five epics implemented.** Parsers
+(PDF/HTML/Markdown/JSON) are real and wired into the live pipeline; chunking
+offers `fixed_tokens` / `sentence_aware` / `heading_aware` with a pluggable
+tokenizer; sources are pluggable (local/S3/HTTP, streaming); runs are idempotent
+with real content-hash deduplication; and operational hardening adds per-document
+error isolation, failure thresholds, an exit-code contract, structured logging,
+and an `ingestion_runs.jsonl` observability hook. See `docs/mvp-ingestion.md`.
+
+**Why:** `llm_knowledge_ingestion` was explicitly an MVP. Parser classes for PDF/HTML/Markdown raised `NotImplementedError`. Chunking was whitespace-window only with a hardcoded guard against any other strategy. The dedup module computed hashes but nothing compared them. Sources were local filesystem only. A single bad file aborted the whole run. This milestone closed those gaps.
 
 ### Ingestion epics