feat: Add DYNAMIC_AGE clamp stats and close-reason max_age field

[keitaj]

Summary

• Adds per-coin clamp counters to the DYNAMIC_AGE summary log so operators can see at a glance whether DYNAMIC_AGE_MIN is the binding constraint for a coin (min=85%) or the value moves freely.
• Adds an optional effective_max_age argument to _record_close and surfaces it as a trailing max_age=Ns on each [close-reason] log line. This lets post-hoc analysis tell apart a long-aged force-close caused by a tight effective_max_age (DYNAMIC_AGE clamped low) from one that ran the full grace window.
• Pure observability — no strategy behavior change. The new field is appended after the existing last_tier= so the close-event regex used downstream still matches.

Changes

• strategies/market_making_strategy.py
  • New _dynamic_age_clamp_stats dict (per-coin: min_clamp, max_clamp, mid, raw_sum, raw_min, raw_max, samples).
  • _get_dynamic_position_age increments the right counter on each compute.
  • _log_dynamic_age emits one [mm] dyn-age <coin> samples=… min=…% mid=…% max=…% raw_avg=…s raw_range=[…-…] line per coin in addition to the existing snapshot. Sorted by min_clamp pct descending. Counters reset after each emit.
• strategies/mm_position_closer.py
  • _record_close gets an optional effective_max_age parameter; emits max_age=Ns suffix when supplied.
  • All 5 call sites pass the value: manage() (unrealized loss path, force-close path) uses the effective_max_age already computed; cleanup_closed and on_position_closed pull it from a new _last_effective_max_age snapshot dict populated each manage() cycle. The dict is evicted on close.
• Tests
  • test_dynamic_age_log.py: new TestDynamicAgeClampStats class covering counter increments for the three clamp regions, summary line format, sort order, and reset behavior. Existing fixture updated to initialize the new dict.
  • test_close_reason_logging.py: new TestRecordCloseEffectiveMaxAge and TestEffectiveMaxAgePropagation covering field presence/absence, regex backward compatibility, and the manage→cleanup_closed propagation path.
  • test_volatility_position_age.py: fixture updated for the new dict.

Test plan

• flake8 (max-line-length 120) passes on the full repo.
• pytest tests/ -q passes (930 tests, +12 new).
• Existing [close-reason] regex last_tier=(\\S+) still captures the same group on the new log format (verified by test_existing_close_reason_regex_still_matches).
• No production behavior change — only logging output is different.

🤖 Generated with Claude Code

[keitaj]

Review: DYNAMIC_AGE clamp stats and close-reason max_age field

Pure-observability change. No strategy behavior is touched; the new field on [close-reason] is appended after last_tier= so the existing last_tier=(\S+) regex still captures the same group (verified by test_existing_close_reason_regex_still_matches). Thread-safety of the new _last_effective_max_age dict mirrors the existing _open_positions pattern (atomic dict ops, no lock needed across main + WS threads). The _log_dynamic_age timer is now advanced after a successful emit instead of before — a minor improvement that means a logger exception leaves the next cycle to retry rather than silently skipping.

---

1. Boundary classification at the clamp endpoints (nit)

strategies/market_making_strategy.py (clamp classification block):

if raw_age <= self._dynamic_age_min:
    stats["min_clamp"] += 1
elif raw_age >= self._dynamic_age_max:
    stats["max_clamp"] += 1
else:
    stats["mid"] += 1

<= and >= mean a raw_age exactly equal to the floor/ceiling counts as a clamp even though no clamping happened. Operationally fine — "value is sitting at the floor" is exactly the signal operators want. Worth keeping in mind only if someone later expects strict-clamp counts (e.g. for a unit-test on a specifically-calibrated baseline). No change requested.

---

2. Defensive .get("samples", 0) (nit)

strategies/market_making_strategy.py (_log_dynamic_age):

samples = int(stats.get("samples", 0))

stats was created via setdefault with all keys including samples, so direct stats["samples"] would never KeyError. Defensive .get is harmless but slightly noisy. Not worth changing.

---

Verdict: LGTM

No blockers, two nits, both ignorable. Existing close-event regex backward-compat is explicitly tested. CI green per author note; will re-run before merge.