feat: Add per-coin unrealized_loss_close_bps overrides

[keitaj]

Summary

• Adds a new coin_unrealized_loss_overrides config key (CLI flag --coin-unrealized-loss-overrides) so the unrealized-loss early-taker-close threshold can be tuned per coin instead of being applied uniformly.
• Mirrors the existing coin_offset_overrides / coin_spread_overrides / coin_size_overrides pattern: same parser, same DEX-prefixed-or-bare lookup, same Dict[str, float] shape on PerCoinOverrides.
• Setting an override to 0 disables the unrealized-loss check for that coin even when the global threshold is non-zero — useful for stable / illiquid coins where the early-close path was firing too eagerly.

Changes

• strategies/mm_config.py: PerCoinOverrides gains unrealized_loss: Dict[str, float] = {}; from_legacy_dict parses coin_unrealized_loss_overrides.
• strategies/mm_position_closer.py: new constructor kwarg coin_unrealized_loss_overrides; new _get_unrealized_loss_bps_for_coin(coin) helper; the manage() early-close block reads through the helper. The warning log now shows the threshold actually applied.
• strategies/market_making_strategy.py: passes the override dict to PositionCloser (alongside the existing coin_spread_overrides); logs the configured overrides at startup.
• bot.py: --coin-unrealized-loss-overrides argparse entry; _STRATEGY_PARAMS['market_making'] includes coin_unrealized_loss_overrides.
• README.md: per-coin overrides section and YAML reference updated.
• tests/test_mm_config.py: PerCoinOverrides defaults assert the new field; full-dict round-trip checks the new key.
• tests/test_unrealized_loss_close.py: new TestCoinUnrealizedLossOverrides covering relaxed override, tightened override, unspecified-coin fallback, DEX-prefixed-vs-bare lookup, the disable-via-zero behavior, and the warning-log threshold reporting.

Test plan

• flake8 passes on the full repo.
• pytest tests/ -q passes (937 tests, +7 new).
• Existing [mm] Position … unrealized loss …bps exceeds threshold …bps … warning text format is preserved (just reports the per-coin value when an override applies).
• No production behavior change when the new flag is absent.

🤖 Generated with Claude Code

[keitaj]

Review: per-coin unrealized_loss_close_bps overrides

Tightly mirrors the existing coin_offset / coin_spread / coin_size override pattern — same parser, same DEX-prefix-or-bare lookup helper layout. The change is small, fully covered by tests (relaxed override, tightened override, unspecified-coin fallback, DEX-prefix→bare resolution, disable-via-zero, and the log-message threshold reporting), and backward-compatible: the new ctor kwarg defaults to None, the legacy-dict key defaults to '', and the gate > 0 keeps the existing "global = 0 means disabled" semantics intact.

---

1. Negative override values are silently treated as disabled (nit)

strategies/mm_position_closer.py (_get_unrealized_loss_bps_for_coin + the > 0 gate in manage()):

If a user passes --coin-unrealized-loss-overrides INTC:-5, the override resolves to -5.0, and the if unrealized_loss_threshold > 0 gate quietly disables the feature for that coin without warning. This matches existing behavior of parse_coin_overrides (no validation of value ranges) and the analogous _get_spread_for_coin paths, so it's consistent with the rest of the codebase. A future hardening pass could centralize "non-negative numeric override" validation in parse_coin_overrides or in each consumer's helper. Not for this PR.

---

2. Log-message threshold reporting (worth noting, not a fix)

strategies/mm_position_closer.py (the warning log inside the early-close branch):

logger.warning(
    f"[mm] Position {coin} unrealized loss {unrealized_bps:.1f}bps "
    f"exceeds threshold {unrealized_loss_threshold}bps -- "
    ...
)

The literal value in the message has changed from self.unrealized_loss_close_bps (global) to unrealized_loss_threshold (effective per-coin). This is the intended behaviour and is more useful for ops, but it's a log-format change — anyone with a downstream parser keyed on the global value would notice. Repo-internal parsers (the [close-reason] regex in /review) only target the structured [close-reason] line, not this warning, so we're fine. Worth keeping in mind if external dashboards exist.

---

Verdict: LGTM

No blockers, no should-fixes, two informational observations. CI green per author note; will re-run before merge.