feat: Auto-exclude coin on consecutive adverse-selection windows

[keitaj]

Summary

• Add an opt-in mechanism that automatically pauses a coin when the AdverseSelectionTracker reports moderate adverse selection (avg_<window> at or below --auto-exclude-threshold-bps, default -3.0) for --auto-exclude-consecutive summary windows in a row (default 3, ~15 min at the default 300s log interval). The coin is paused for --auto-exclude-cooldown seconds (default 1800) and then automatically resumes.
• Complements two existing defenses: loss_streak_limit (close-PnL based, weak against slow bleed) and dynamic_offset (caps at +3 bps offset addition). Auto-exclude reacts to the mid-price drift signal directly and stops a coin that is structurally adverse-selected.
• Builds on the dataclass scaffolding from refactor: Group MM config into dataclasses (Phase 1) #128 / refactor: Group MM config into dataclasses (Phase 2a) #129 / refactor: Group MM config into dataclasses (Phase 3) #131 — uses a new AutoExcludeConfig group on MMConfig, with the legacy flat keys still accepted via from_legacy_dict and CLI flags wired with dest=<config_key> per refactor: Align argparse dest with config keys (Phase 2b) #130.

Why

AdverseSelectionTracker already produces a per-coin [adverse] Summary log every 300s. Until now the markout signal was observed only by humans (or by dynamic_offset's short 5s window). When a coin shifts to a structurally adverse regime mid-session, the existing guards do not stop it; this PR closes that gap with a small, self-contained feedback loop.

Changes

• strategies/mm_config.py: new AutoExcludeConfig dataclass with __post_init__ validation (consecutive ≥ 1, min_fills ≥ 1, cooldown > 0, window_label ∈ {5s,30s,60s}). Added to MMConfig and from_legacy_dict.
• ws/adverse_selection_tracker.py: each _log_summary window also appends a snapshot (ts, fills, avg_5s/30s/60s) to a per-coin deque (cap MAX_HISTORY = 10). New get_recent_windows(coin, n) returns the last n entries.
• strategies/market_making_strategy.py: new _check_auto_exclude(coin) invoked from run() just before the cooldown skip. Defensive getattr(self, 'cfg', None) so tests that bypass __init__ continue to work. The cooldown deadline check in run() is no longer gated on loss_streak_limit > 0, since auto-exclude alone can now drive cooldowns through the shared _coin_cooldown_until map.
• bot.py: six new CLI flags (--auto-exclude, --auto-exclude-threshold-bps, --auto-exclude-consecutive, --auto-exclude-min-fills, --auto-exclude-cooldown, --auto-exclude-window-label) and corresponding entries in _STRATEGY_PARAMS['market_making'].
• README.md: human description in the market_making section + AI YAML reference entries.

Backward compatibility

• Default-disabled. Existing deployments are unaffected unless they pass --auto-exclude together with --enable-adverse-selection-log.
• No changes to risk_manager.py Level 1 guardrails.
• Existing flat instance attributes preserved as before; no new internal API breakage.

Test plan

• Unit tests added:
  • tests/test_mm_config.py: dataclass defaults + four validation cases + extended from_legacy_dict coverage.
  • tests/test_adverse_selection_tracker.py: append-on-summary, get_recent_windows last-n / unknown coin / n<=0 / MAX_HISTORY cap / avg=None semantics.
  • tests/test_adverse_auto_exclude.py (new): trigger / borderline / threshold-exact / short history / min_fills / None-avg / 30s window label / cooldown lifecycle / shared cooldown coexistence / constructor validation propagation.
• flake8 passes.
• pytest passes (879 tests total, +28 new).
• No regressions in existing tests.

🤖 Generated with Claude Code

[keitaj]

Review: feat: Auto-exclude coin on consecutive adverse-selection windows

Self-contained feature that uses existing AdverseSelectionTracker infrastructure. Validation, history capping, cooldown sharing with loss_streak, and the test matrix all line up with the design doc. Default-disabled, so deployment risk is minimal.

A couple of nits, no blockers.

---

1. Reacquiring the lock per coin in _log_summary (nit)

ws/adverse_selection_tracker.py L222-223:

for coin in sorted(aggregates.keys()):
    ...
    with self._lock:
        self._history[coin].append(snapshot)

The lock is acquired once per coin inside the loop. With ~10–20 coins per summary the overhead is negligible, but the pattern is slightly off — every other path in the file batches under a single with self._lock:. Hoisting the lock outside the loop both reads better and is consistent.

Suggestion: Move the with self._lock: to wrap the whole for coin in sorted(...) block, or build a list of (coin, snapshot) pairs first and append them all under one lock.

---

2. Cooldown-expired log doesnt identify which feature triggered it (nit)

strategies/market_making_strategy.py L401-407:

When the deadline is reached the message is "[mm] {coin} cooldown expired, resuming" — fine, but with two writers to _coin_cooldown_until (loss_streak and auto_exclude), its no longer obvious from the log which feature paused the coin in the first place. The trigger-time log already names the source (hit N consecutive losses vs auto-excluded: ...), so this only matters for the resume side.

Suggestion: Optional. Track the trigger source alongside the deadline (e.g. _coin_cooldown_source: Dict[str, str]) and include it in the resume log. Easy follow-up if it ever matters in the field.

---

3. Disabled-path tests do not assert the tracker was untouched (nit)

tests/test_adverse_auto_exclude.py TestDisabled:

The two disabled-feature tests check that no cooldown was set, but dont verify that tracker.get_recent_windows wasnt called at all. Strengthening to tracker.get_recent_windows.assert_not_called() would catch a future refactor that accidentally moves the disable check below the tracker query.

Suggestion: Optional. Add the assertion in test_disabled_does_nothing.

---

Other observations (no action needed)

• Cooldown skip ungating: removing if self.loss_streak_limit > 0: around the cooldown deadline check is correct now that auto_exclude also writes to the same map. Existing loss_streak callers are preserved because the inner reset (_loss_streaks[coin] = 0) is now guarded on loss_streak_limit > 0. ✅
• getattr(self, cfg, None) defensive guard: matches the bypass-__init__ test pattern used elsewhere in the suite. ✅
• Threshold semantics: avg > threshold early-returns, so avg == threshold triggers — the exact-match test pins this down explicitly. ✅
• Re-trigger after cooldown expires: by design, if the same N adverse windows are still in the deque after the cooldown elapses, the coin will be re-excluded on the next _check_auto_exclude. This matches the "structurally adverse coin" use case and is consistent with the design doc. ✅
• History cap test: the MAX_HISTORY + 3 scenario verifies FIFO eviction with the exact expected oldest/newest values — strong assertion that catches real regressions. ✅

---

Verdict: LGTM

Three nits, all optional follow-ups. Phase 4 of the MM hardening arc is in good shape — merging on CI confirmation.