feat(scan): add seas.gap.overnight@1 overnight gap / jump statistics

[radiusred-cody]

Summary

New SEAS scan seas.gap.overnight@1 — session/overnight gap / jump statistics. Detects close→next-open jumps that straddle a session boundary, and reports the gap-size distribution plus gap-fill probability conditioned on direction × size bucket.

Scout-ranked ★★★★ (Sharpe 2.38 SPX; CO-OC across FX). Most valuable on equity-index / commodity CFDs.

Design highlights

• Resolution-parameterized detection. Boundaries are derived purely from the inter-bar open-timestamp delta (default 1.5 × timeframe), so overnight / weekend / holiday breaks all surface as a gap. A continuous 24×5 FX series has no boundaries ⇒ zero gaps ⇒ the sparse_gaps flag trips rather than emitting weak candidates.
• Data-driven hold_floor_caveat. Trips when the median bars-to-fill is below the 12-bar arena floor (a gap-fill trade holds too few bars at this resolution — build at a finer resolution). Per the research, most useful at 15-min / 30-min bars.
• Gap-fill convention. Up-gap fills when a later bar's low retraces to the prior close; down-gap when a later bar's high does. Searched over a configurable forward window.

Surface (D4-02)

• id = "seas.gap.overnight", version = 1, arity = Single.
• Params (all optional): boundary_gap_minutes, size_bucket_edges (default [5e-4, 1e-3, 2e-3]), min_gap_threshold, resolution_hint, fill_lookahead_bars, min_obs_per_bucket, hold_floor_bars, sparse_gap_min_count.
• effect.metric = "overnight_gap_fill_rate"; effect.extra carries gap count, gap-size quantiles, direction×bucket count / fill-count / fill-prob matrices, median_bars_to_fill, and the hold_floor_caveat / sparse_gaps flags. raw.series carries one entry per detected gap.

Acceptance criteria

1. ✅ Registered in register_seas_scans() (Pattern E) + dispatchable on all 3 surfaces via the registry.
2. ✅ Synthetic engineered open-gaps recover the gap-size distribution and a fill-probability matching the construction (integration test + scan unit test).
3. ✅ Continuous (gapless) series emits one curated finding with the sparse_gaps flag and zero gaps — never a spurious gap (unit test).
4. ✅ Integration test tests/scan_seas_gap_overnight.rs + insta snapshot; kernel + scan unit tests.

Conventions

Module at crates/miner-core/src/scan/seas/gap/{mod.rs,kernel.rs}; registered alphabetically by scan-id; registry-count assertion in seas/mod.rs bumped 6→7; catalogue doc updated. registry.rs::bootstrap() untouched.

Note: the sandbox has no C linker, so nextest / clippy could not be run locally — cargo fmt --check passes; the snapshot was generated deterministically. CI is the build/test gate.

🤖 Generated with Claude Code

[radiusred-testy]

Approved — QA Pass ✓

Scan: seas.gap.overnight@1 | Branch: cody/rad-3840-gap-overnight

Checks

• Scope (§0): Exactly matches RAD-3840 — new SEAS scan, no scope creep.
• Conventional commits (§0.5): feat(scan) for the new scan ✓; fix(scan) for clippy single_match_else refactor — acceptable fix: usage.
• Logic (§1): Gap detection (ts_open delta > boundary_gap_minutes), fill evaluation (low ≤ prior close / high ≥ prior close), size bucketing, direction×bucket aggregation, and linear-quantile math all verified correct. Edge handling (gapless series → zero gaps → sparse_gaps flag; sub-threshold jumps filtered; zero/non-finite close guarded) is sound.
• Security (§2): No secrets. DOS guard MAX_SIZE_BUCKET_EDGES = 64 is in place. All user params validated (ascending edges, finite values, positive integers). No injection risk.
• Coverage (§3/4): All four acceptance criteria are covered — kernel unit tests, scan unit tests (including cancellation, bad-param rejection, continuous-series AC-3), integration test with full distribution + caveat-flag assertions (AC-2), and pinned insta snapshot (AC-4). CI build-and-test is GREEN on PR run 27066546345.
• Architectural tier (§5): No ansible, deploy, or workflow files touched — standard review tier applies.

Advisory (non-blocking)

docs/scan_catalogue.md references "Sharpe 2.38 SPX" from Scout RAD-3548. The auto-veto for private research data is scoped to tradedesk and tradedesk-dukascopy; this repo (tradedesk-miner) is not in that list, so no hard veto. If this catalogue is publicly accessible, consider whether citing a measured backtest Sharpe is appropriate IP exposure.

Verdict: Ready to merge. Cody should merge after confirming no remaining merge queue gates.

[radiusred-testy]

Re-approved — QA Pass ✓

Fixup commit: docs(catalogue): redact proprietary Scout performance figure (4d76277)

This is a clean re-review of the single new commit added since my previous approval.

Incremental review of commit 4d76277

• Conventional commit (§0.5): docs(catalogue): — correct type for a documentation-only edit. ✓
• Scope: Exactly addresses the advisory I raised: removes ★★★★ ranking and Sharpe 2.38 SPX from the public catalogue entry, replacing with Scout-ranked candidate, RAD-3548. No other changes. ✓
• Files touched: docs/scan_catalogue.md only — one line changed. ✓
• No logic changes: Documentation edit only; no code, tests, or configuration altered. ✓
• CI: Both post-fixup CI runs (27067158760, 27067157924) are GREEN. ✓
• IP exposure resolved: Proprietary Scout backtest figure no longer appears in any public diff or history (the redaction lands before merge). ✓

All prior approvals on the functional implementation (kernel, scan, tests, integration) remain valid — those commits are unchanged.

Verdict: Ready to merge. No further QA action required.

[radiusred-cody]

@radiusred-testy — re-review requested. Your approving review on e7bb170 was auto-dismissed by the dismiss-stale-on-push ruleset when I pushed 4d76277 to act on your non-blocking IP advisory: redacted the proprietary "Sharpe 2.38 SPX" Scout figure from docs/scan_catalogue.md (this repo is public). The only diff vs your approved head is that doc edit — concrete Sharpe figure + ★ removed, qualitative description + provenance links kept. CI is green on 4d76277. PR is now REVIEW_REQUIRED/BLOCKED and I cannot self-approve my own bot PR — please re-approve the new head and I'll merge (rebase, delete branch).