Add research autolearn, counterfactual twin, and deep backfill tooling

[ArielB1980]

Summary

• add continuous research autolearn/campaign-gate/filter/funnel tooling and wire richer replay/live gate instrumentation to make blocker diagnosis deterministic
• add counterfactual twin utilities and tests for decision-tape uplift comparison and batch ranking of candidates
• add CoinAPI-backed deep backfill path plus related data acquisition support to improve research candle completeness
• extend targeted unit coverage for research evaluator eligibility/timeout behavior and per-symbol config resolver behavior

Test plan

• uv run pytest tests/unit/test_research_filter_report.py tests/unit/test_research_autolearn.py tests/unit/test_research_campaign_gate.py tests/unit/test_research_evaluator_eligibility.py tests/unit/test_research_evaluator_replay_timeout.py tests/unit/test_counterfactual_twin.py
• reviewed git diff --stat main...HEAD for scoped changes
• run wider integration suite before merge (recommended)

Made with Cursor

---

Note

Medium Risk
Medium risk because it touches core trading/replay pathways (LiveTrading, ExecutionGateway, replay exchange sim) with new replay-mode bypasses and logging/instrumentation that could affect behavior if accidentally enabled in non-replay environments.

Overview
Adds a new continuous research toolchain: scripts to run campaign-level stop gating, auto-generate next-cycle env overrides from prior runs, summarize ablation batches, and produce blocker/funnel reports from run logs and recorded decision events.

Introduces counterfactual decision-tape instrumentation by emitting COUNTERFACTUAL_DECISION/COUNTERFACTUAL_ACTION events with stable decision_ids, plus a new counterfactual_twin module to stitch these events into a deterministic tape and score parameter overrides via utility-uplift ranking.

Expands the replay/research harness for higher determinism and throughput: symbol resolution across spot/futures naming, explicit order-update polling, forced flatten-at-end, optional DB mocking, relaxed replay gates/limits via env flags, and additional replay metrics (e.g., paused ticks, trades closed).

Adds CoinAPI-backed deep backfill support: a new CoinAPIClient, DataAcquisition.fetch_spot_historical(..., source=...) routing, and a deep_backfill_daily.py utility for large lookback candle ingestion; also updates research control/nightly scripts to default to replay-oriented settings and add continuous-daemon commands.

^{Written by Cursor Bugbot for commit 31003e9. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 4 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Replay bypass skips kill switch return, allowing continued tick processing

High Severity

When replay_fail_open is True (DATA_FAILURE kill switch in replay mode), the code logs a warning but does not return. Execution falls through past the entire kill switch block into normal tick processing (order timeouts, data health checks, analysis, etc.). The original code always returned after handling an active kill switch. For non-DATA_FAILURE kill switch reasons in replay mode, the else branch handles them and returns — but the DATA_FAILURE path silently continues. This means replay ticks proceed with an active kill switch, which contradicts the safety invariant that no new entries occur while the kill switch is active.

 

[cursor]

NO_SPEC bypass proceeds without spec causing downstream failures

Medium Severity

When _replay_relaxed_signal_gates is True and no instrument spec is found, the code logs and continues instead of skipping. However, spec remains None and is used downstream for step size calculation, min-size enforcement, and position sizing. Code later references spec.size_step, spec.tick_size, etc., which will raise AttributeError on None. The bypass only logs — it doesn't provide a fallback spec object.

 

[cursor]

Symbol resolution cache returns empty string as cache miss

Low Severity

_resolve_market_symbol checks if cached: which treats an empty string as a cache miss. While the early return for empty raw prevents empty strings from being cached, if a symbol resolves to a non-empty string that's falsy in some edge case, the cache would be bypassed. More importantly, if any resolved symbol happens to be a string that Python considers falsy, the cache lookup would fail and re-resolve every time, causing a performance issue.

 

[cursor]

Counterfactual twin sort order inverts candidate ranking

Medium Severity

evaluate_candidate_batch sorts by a tuple including -int(r["candidate_open_count"]) as the third key with reverse=True. The negative sign combined with reverse sort means candidates with more opens rank lower (double negation). If the intent is to rank higher-open-count candidates higher when uplift and delta are tied, the negation produces the opposite order.