docs: add custom data loader guide

[mvanhorn]

Summary

Adds agent/backtest/loaders/README.md, a guide for registering a custom
historical OHLCV data loader, and a one-line pointer to it from
CONTRIBUTING.md. It documents the loader contract end to end: the
DataLoaderProtocol shape, the @register decorator, the registry / config
wiring, and how to select the loader via source="<your loader>". This helps
contributors who want to feed a data source the project does not ship a
connector for into the backtest data layer.

Why

Issue #178 asks how to feed a custom data source into the backtest data layer.
The registration recipe currently lives only in a maintainer reply, not in the
repo — so a new contributor has no in-tree path to follow. This guide closes
that gap and also documents two steps that are easy to miss: adding the module
to the lazy import list so @register fires, and adding the source name to the
config schema's allowlist so config.json validation accepts it.

Changes

• New agent/backtest/loaders/README.md:
  • The loader contract (name / markets / requires_auth / is_available()
    / fetch() returning {symbol: DataFrame} with a trade_date index and
    open/high/low/close/volume columns).
  • A 5-step quickstart: create the module, add it to _loader_modules in
    registry.py, add the name to _VALID_SOURCES in runner.py, optionally
    add to FALLBACK_CHAINS, and select via source=.
  • An explicit "out of scope: real-time data" section pointing at the broker
    connectors, plus a checklist.
• Edit CONTRIBUTING.md: a short "Adding a Custom Data Loader" section
  linking to the new README.

Deliberately out of scope: real-time tick/depth ingestion. There is no
first-class interface for pushing a custom real-time stream today (real-time
flows through the broker connectors), so the guide documents the historical path
only and points readers there.

Test Plan

Documentation only — no code paths change. Every symbol, signature, and path in
the guide was verified against the current source:

• agent/backtest/loaders/base.py — DataLoaderProtocol and the fetch
  signature / OHLCV return contract; check_budget, retry_with_budget,
  cached_loader_fetch, loader_cache_get/put, validate_date_range.
• agent/backtest/loaders/registry.py — register, the _loader_modules list
  in _ensure_registered(), FALLBACK_CHAINS, get_loader_cls_with_fallback.
• agent/backtest/loaders/yfinance_loader.py, okx.py, akshare_loader.py —
  the concrete loader pattern the examples mirror.
• agent/backtest/runner.py — BacktestConfigSchema._VALID_SOURCES /
  valid_source (the config-validation step) and _get_loader.

No live, broker, MCP, network, secret, or deployment surface is touched.
Rollback: revert the single commit with git revert.

Checklist

• Docs-only change; no code paths modified.
• Every symbol/path in the guide verified against the current source files.
• Commit signed off (git commit -s) per the DCO requirement.

Fixes #178

[warren618]

Thanks @mvanhorn — really clean PR, and I appreciate you verifying every symbol against the source.

One change on the way in: instead of a standalone agent/backtest/loaders/README.md, we folded this into the Detailed Capabilities → Custom Data Source section of the READMEs (all 5 locales), so the people who actually hit this — users wiring up their own data source — find it right next to the existing data-source docs instead of in a separate file. I pushed that reshape to your branch and kept your authorship and sign-off intact. Substance is unchanged; I only:

• trimmed the parts that restated DataLoaderProtocol (it already lives in base.py, now pointed to as the source of truth),
• switched the example to the duck-typed shape (@register on a class that satisfies the protocol — no base class to inherit), and
• noted the steps edit package source, so it's a from-clone (pip install -e .) workflow.

Merging now — thanks for closing the gap on #178! 🙏

[mvanhorn]

Thanks for merging the loader guide too, @warren618. Custom data loaders are a lot more approachable documented.