v0.1.2 — first matrix-tested release

0.1.2 — 2026-05-04

Second public release. Hardens the dependency-injection contract, trims the
CLI surface, fixes a backlog of path-resolution bugs surfaced by host-app
integration, and ships llms.txt directly into scaffolded workspaces.

Breaking changes

• PathsConfig is now required everywhere. Every library entry point
  (run_data_pipeline, run_backtest, run_best_trial,
  run_indicator_calculation, WFARunner, PortfolioBacktestRunner, …)
  takes paths: PathsConfig as a kwarg-only required parameter. Library
  code no longer falls back to PathsConfig.from_env(); the strategy
  bridge raises CFG-003 if path strategy params aren't injected.
  Construct one PathsConfig at startup (via
  PathsConfig.from_project_root(<root>) or from_platformdirs) and
  thread it through.
• echolon run removed. Use echolon backtest single <strategy_dir> —
  it walks up to the workspace marker and recovers ctx, so no flags are
  required.
• echolon init-strategy back-compat alias removed. Use echolon init.
• Default workspace layout flattened: workspace/strategy/baseline/ +
  workspace/backtest/ (was workspace/current/...). Host apps with
  iteration-loop layouts override via the workspace marker's paths field.
• Legacy strategy imports: from ...core.base.X no longer resolves.
  Run echolon migrate <strategy_dir> to rewrite to the current absolute
  paths (echolon.strategy.X).
• Position field cleanup: entry_price, side, is_short,
  is_flat removed; only is_long survives. Use position.avg_price
  for the entry price; check direction via position.direction
  ("LONG" / "SHORT" / "FLAT").
• Pydantic v2 idioms: strategy output schemas use
  model_config = ConfigDict(extra="allow") instead of the deprecated
  class Config: block.

New

• llms.txt shipped into every workspace. echolon init /
  echolon hello drop the agent orientation manual at the workspace
  root, so an LLM agent walking into the project finds it without
  needing the MCP server up.
• Data pipeline auto-populates main_contract.csv. Step 6 of
  run_data_pipeline either copies a curated main_contract.csv from
  the raw data tree or derives it via the max-volume rule. Previously a
  separate scaffolding step.

Fixed

• WFA OOS replay: paths threaded through BacktestRunner.best_trial,
  get_trading_calendar_instance, load_indicator_metadata, and the
  SHFE adapter's market_data_dir. Resolves a chain of CFG-003
  errors that fired on every per-window backtest.
• Contract-aware broker now probes slot-style indicator dir first and
  falls back to instrument-style
  ({indicator_dir}/{slot}/by_contract/ or
  {indicator_dir}/{instrument}/by_contract/) — supports both
  echolon backtest single (slot layout) and host-app iteration-loop
  pipelines (instrument layout).
• SHFE day-data extractor default subdir aligned with the natural
  convention: {raw_data_dir}/SHFE/day_data/ (was the misleading
  raw_data/).
• README runtime-registration table: corrected OpenAI Agents SDK row
  (previous form raised TypeError; now uses the required params
  dict), added the missing outer-wrapper note for Cursor's mcpServers
  config, replaced the vague Codex CLI pointer with the documented
  codex mcp add command, replaced the LangChain pointer with a
  copy-pasteable MultiServerMCPClient snippet.

Internal cleanup

• Removed validation-backup/ skill tree from echolon (relocated to
  qorka — the KEEP/REVERT workflow is qorka-internal, not generic
  backtest infra).
• Renamed test fixture al_v6_1_migrated/ → aluminum_baseline/ and
  scrubbed qorka-specific path references from fixture file headers.
• Skill-doc audit pass: removed qorka-internal class names
  (ExplorationOrchestrator / ExploitationOrchestrator),
  internal-task-tracker references ("Task 6 audit"), and outdated
  migration framing.
• README rewrite (EN + zh-CN): trimmed from ~265 lines to ~104,
  restructured around the two journeys readers actually have (try via
  CLI in 30 seconds; drive from an LLM agent). Native-Chinese tone
  pass to drop translation-shaped phrasings.