Release Notes — Sirchmunk v0.2.0

Features

LENS DEEP engine: prior-warmed agentic loop (PWAL) unifying retrieval + synthesis, belief-driven action loop, and rga match-line anchoring.
Multi-path retrieval + confidence fusion: parallel lexical / exact-entity / directory / structure / cross-document topic-map routes fused via confidence-weighted RRF, with soft route-collapse fast-tracking.
Format & structure coverage: native LOG/PPTX/XLSX exact-match fallback, heuristic document-tree v2 (incl. DOCX/RST), and default structure anchors.
Grounded numeric verification: computation answers re-checked deterministically from model-disclosed, evidence-grounded operands.
Hard token budget: task-local token reservation with dynamic evidence preloading.
ResearchOps benchmarking: lifecycle / scaling / Pareto evaluation, control gates, and a full HotpotQA suite (sampling governance, raw-corpus sync, dynamic corpus).
Baselines: one unified retrieval contract, a closed-book arm, plus Hybrid RAG and LightRAG v1.3.6 lifecycle baselines.
Semantic judge: surface-form-tolerant EM with calibrated, blind-agreement-measured judging.
Web: interactive knowledge-graph / cluster visualization with i18n.
Model support: rich responses, resilient retrieval, and a DeepSeek V4 thinking-mode switch.

Large-corpus stability: bounded per-file/query retrieval cost (rga adapter whitelist, size cap, tiered rg-first, per-file match cap) + directory scan on by default — eliminates the archive/large-tree rga timeouts and "no results" failures.
Rescue loop-path refusals instead of abstaining silently; recover garbage output and calibrate answer granularity.
Restore paginated-document handling in PWAL; harden _sanitize_answer_output.
Validate stale-index; accept nested evidence in the retrieval contract.
Fix a shared-search concurrency issue; make HOTPOT_MAX_CONCURRENT reach the baseline suite; fix benchmark timestamps and quickstart limits.

Overlap DEEP query analysis with retrieval probes to reduce latency.
Harden baseline cache identity; improve HotpotQA quickstart reporting, diagnostics, and guardrails; allow per-baseline sample concurrency.

Replace benchmark-specific hard rules with generalizable logic (grounded computation-trace; corpus-DF adaptive stop-words + injectable tokenizer).
Make abstention a rendering decision, not a synthesis branch; move benchmark answer semantics out of the retrieval chain.
Remove the FinanceBench benchmark in favor of the HotpotQA ResearchOps suite.

Rewrite AGENTS.md in English (RFC 2119) with an anti-hardcoding / anti-overfitting policy and retrieval cost invariants.
Add v0.1.0 News; refresh README architecture figures and knowledge-graph screenshot; sync the LENS paper; English benchmark README; document ResearchOps usage and env templates.

fix: use DeepSeek V4 thinking switch by @chenshibin-vip in #178
LENS DEEP Retrieval & ResearchOps Enhancement by @wangxingjun778 in #179
[Feature & Fix] multi-path retrieval, large-corpus robustness, and ResearchOps benchmarking by @wangxingjun778 in #180
update news by @wangxingjun778 in #181
update readme and version by @wangxingjun778 in #182

Full Changelog: v0.1.0...v0.2.0