refactor: Group MM config into dataclasses (Phase 3)

[keitaj]

Summary

• Final phase of the MM config refactor (after refactor: Group MM config into dataclasses (Phase 1) #128 / refactor: Group MM config into dataclasses (Phase 2a) #129 / refactor: Align argparse dest with config keys (Phase 2b) #130). Two more grouped dataclasses cover the remaining flat config keys:
  • DynamicOffsetConfig — auto-adjusted BBO offset based on observed adverse selection (7 keys).
  • DynamicAgeConfig — volatility-adjusted MAX_POSITION_AGE (4 keys).
• MarketMakingStrategy.__init__ reads these groups from self.cfg.*. The existing flat instance attributes (self._dynamic_offset_enabled, self._dynamic_age_max, etc.) are preserved as aliases so internal references and tests are unchanged.
• Also refresh the MMConfig docstring — it had been stale since Phase 1 listing close/imbalance/schedule as "future phases".

Why

After this PR, every CLI/env-driven MM parameter that has a stable home flows through MMConfig. The remaining config.get(...) calls in MarketMakingStrategy.__init__ are limited to the core MM primitives (spread_bps, order_size_usd, bbo_offset_bps, etc.) and the parameters passed straight through to PositionCloser (max_position_age_seconds, taker_fallback_age_seconds, aggressive_loss_bps).

This closes out the dataclass-grouping work that #128 began. CLI surface, env vars, and existing test suite are unchanged.

Changes

• strategies/mm_config.py: add DynamicOffsetConfig and DynamicAgeConfig dataclasses; extend MMConfig and from_legacy_dict() with the two groups; refresh MMConfig docstring.
• strategies/market_making_strategy.py: read dynamic-offset and dynamic-age parameters from self.cfg.*. Existing flat instance attributes preserved as aliases.
• tests/test_mm_config.py: add unit tests for the two new dataclasses and extend from_legacy_dict end-to-end coverage to all the new keys.

Backward compatibility

• CLI flags and environment variables: unchanged.
• Flat instance attributes (self._dynamic_offset_*, self._dynamic_age_*) preserved as aliases of the new cfg fields.
• No new validation introduced; defaults match the legacy config.get() defaults exactly.

Test plan

• Unit tests added (2 new dataclass tests + extended from_legacy_dict coverage).
• flake8 passes.
• pytest passes (853 tests total).
• No regressions in existing tests.

🤖 Generated with Claude Code

[keitaj]

Review: refactor: Group MM config into dataclasses (Phase 3)

Closes out the dataclass-grouping work cleanly. Two more groups (DynamicOffsetConfig, DynamicAgeConfig), the alias pattern stays consistent with #128/#129, and the previously stale MMConfig docstring is now accurate.

Verified:

• All 7 dynamic_offset_* legacy keys map to the matching dataclass field with identical defaults.
• All 4 dynamic_age_* legacy keys map correctly. The dataclass uses more descriptive field names (baseline_vol_bps, min_seconds, max_seconds) while from_legacy_dict reads the original key names — no env / CLI breakage.
• _dynamic_age_log_interval and _base_max_position_age are intentionally left outside the new dataclass: the former is a constified pseudo-config (Phase 1 scope), the latter is core MM plumbing shared with PositionCloser.
• 853 tests pass; flat instance aliases in MarketMakingStrategy keep call-sites and tests untouched.

No blockers, no should-fix.

---

1. No invariant check between min_seconds and max_seconds (nit)

strategies/mm_config.py DynamicAgeConfig:

The dataclass accepts any pair of positive numbers, including min_seconds > max_seconds, which would silently produce a never-satisfied range. The original code didnt validate this either, so this matches existing behaviour — calling it out only because new dataclasses are a natural place to add such guards if you ever want them.

Suggestion: Optional. If desired, add a __post_init__ that raises when min_seconds > max_seconds or when either is non-positive. Defer to a follow-up if youd rather keep this PR purely structural.

---

Verdict: LGTM

After Phase 3 the MM strategys __init__ only retains direct config.get(...) calls for core MM parameters (spread_bps, order_size_usd, etc.) and a few values still flowing straight to PositionCloser. The dataclass-grouping arc started in #128 is complete. Merging on CI confirmation.