feat(evaluation): add full-A strategy and ML OOS ablation

[Atlas2005]

Summary

Adds the PR #121 Full-A strategy and ML out-of-sample ablation workflow for the offline A-share snapshot.

This PR:

• compares fixed strategy candidates, LightGBM ranking, and a validation-only dynamic selector
• uses common chronological train/validation/test folds
• reuses the existing account-aware execution, fill, lot-size, capital, and transaction-cost paths
• validates cost sensitivity at 1.0x, 1.5x, and 2.0x
• preserves purge/embargo based on the target horizon
• records compact ML ranking authenticity evidence
• remains manual-only and offline after snapshot creation
• performs no broker, DeepSeek, credential, or live-trading access

Full-A validation

Snapshot:

• provider: all_a_share_snapshot
• snapshot digest: ca0448aefb17cd2990fd2a095835d6dc44e25046c259748545051516bf303615
• requested symbols: 5,481
• valid symbols: 5,481
• skipped symbols: 0
• common folds: 5
• candidates: 11
• cost-sensitivity rows: 33
• external calls: false

Final V3 run:

• process return code: 0
• timed out: false
• run status: completed
• successful ML folds: 5
• authentic ML folds: 5

Main result

Recommended candidate:

equal_weight_baseline

Metrics:

• compounded return: 36.4493%
• benchmark return: 31.8605%
• compounded excess return: 2.8476%
• positive-return fold ratio: 80%
• positive-excess fold ratio: 60%
• maximum drawdown: -8.9672%
• filled trades: 50
• rejected trades: 0

ML result

The corrected LightGBM path genuinely controlled order selection in all five folds.

Evidence includes:

• nonconstant prediction distributions
• (test_start, symbol) prediction association
• ML-selected symbols passed into the existing executor
• actual filled symbols matching ML-selected symbols
• no silent baseline fallback
• common test ranges and benchmark values matching fixed-rule evaluation

Final LightGBM metrics:

• compounded return: 8.2139%
• benchmark return: 31.8605%
• compounded excess return: -19.5798%
• positive-return fold ratio: 40%
• positive-excess fold ratio: 20%
• maximum drawdown: -7.8151%

This means the ML execution path is valid, but the current feature, label, model, and portfolio mapping do not demonstrate incremental value.

Dynamic selector result

• compounded return: 16.8098%
• compounded excess return: -13.9608%
• uses test outcomes: false
• decisions are based only on prior validation evidence

The selector did not outperform the equal-weight baseline.

Integrity fixes included

• optimized buy-and-hold benchmark construction and added run-scoped caching
• aligned ML and fixed-rule test boundaries and benchmark basis
• propagated actual ML filled/rejected trades, turnover, costs, and execution windows
• fixed ML prediction key and rebalance-date handoff
• rejected missing, nonfinite, constant, or tied ML ranking inputs instead of silently falling back to canonical ordering
• added compact ranking evidence and progress reporting
• corrected disabled dynamic-selector CLI output

Validation

• py_compile: passed
• focused PR feat(evaluation): add full-A strategy and ML OOS ablation #121 tests: passed
• ML factor training tests: passed
• ML robustness tests: passed
• real-data scale-up tests: passed
• announcement tests: passed
• full suite: 1598 passed
• git diff --check: passed

Safety and scope

• no commit generated by automation
• no external API calls during the Full-A validation
• no DeepSeek call
• no broker access
• no credentials accessed
• no live or simulated broker orders
• no branch switch