v0.2.0 — ChaosSMTP, ChaosBlob, engine/MCP fixes

Two new chaos plugins, three engine/MCP fixes, one breaking default change.

See CHANGELOG.md for the full notes.

Highlights

Added

• ChaosBlob — S3-compatible-ish object-storage chaos server (PUT/GET/HEAD/DELETE/ListObjectsV2, blob-specific fault injection, metrics, CLI, presets, fixture, docs).
• ChaosSMTP — outbound-email chaos server. Stage-aware error injection across MAIL/RCPT/DATA, slow-reply latency, accepted-but-dropped messages, three capture modes (metadata/discard/full), admin app, immutable captured-header policy, metrics, CLI, presets, fixture, docs. Mail is never relayed.

Changed (breaking)

• server.workers default is now 1 (was 4). Multi-worker mode is now opt-in and requires a file-backed metrics.database; the previous default silently fragmented metrics across worker processes. Top-level config validators now reject workers > 1 combined with an in-memory metrics database.

Fixed

• Engine: explicit MalformedContentLength rejection (closes the silent size-guard bypass caused by RequestBodyTooLarge subclassing ValueError).
• Engine: graceful config-handoff fallback when the temp file is gone (shared engine/config_handoff.py helper used by blob/llm/web).
• LLM MCP analyze_latency: percentiles now computed over the full population (was the smallest-100 lower tail).
• LLM completions: non-object JSON body returns 400 invalid_request_error.
• Admin /admin/config POST: unknown sections and non-object values rejected with 400.
• LLM X-Fake-Template: hardened against helper errors; random_words(...) hard-capped to 10 000 words.
• MCP SQL query(): row cap enforced via fetchmany, defeating LIMIT -1 bypasses.
• MCP analyzer: now truly read-only (file:...?mode=ro, no WAL/SHM side files).

Removed (security)

• Pre-authorised rm -rf $REPO/.git && git init permission entry from .claude/settings.local.json.