fix: make the bandit reward's ARR term live and warn on missing qlib metrics

[michaelraziel-bit]

Fixes #1451

What was wrong

extract_metrics_from_experiment in rdagent/scenarios/qlib/proposal/bandit.py looked up the annualized return with a trailing space in the key ("...annualized_return "). Qlib never logs that key, .get silently returned the 0.0 default, and the ARR feature was always zero. The eighth feature (arr / -mdd) is derived from it, so it was always zero too. Between them that is 0.45 of the reward weight, so every default rdagent fin_quant run has been steering the factor/model bandit on IC, ICIR, Rank IC, Rank ICIR, IR and drawdown only.

slot	feature	weight	before	after
4	ARR	0.25	always 0	live
7	ARR / |MDD|	0.20	always 0	live

Nothing noticed because every lookup defaulted to 0.0 inside a bare except Exception, so a mistyped key, a missing key and a genuinely zero metric were indistinguishable.

What this PR does

Four commits, each droppable on its own:

1. fix: remove the trailing space. This is the whole bug fix.
2. refactor: move the Qlib metric key strings into rdagent/scenarios/qlib/metrics.py and import them from both bandit.py and feedback.py (IMPORTANT_METRICS is now built from the constants), so the two places cannot disagree on spelling again. Replace the silent defaults with a helper that logs a warning naming the missing key and the keys that were present, then still returns the default so a partial Qlib result degrades instead of crashing the loop. Narrow the outer except Exception to the case it was guarding (experiment.result is None after a failed run). Change the max_drawdown default from 1.0 to 0.0: Qlib reports drawdown as a number <= 0, so 1.0 would have put -1.0 in the feature vector and flipped the sign of the ratio whenever the key was missing.
3. refactor: rename the sharpe field to calmar. The formula (arr / -mdd) is unchanged; the name now says what it computes. Metrics is only constructed inside bandit.py, so nothing else is touched. This commit is optional. If you would rather keep the name, or replace the feature with a true Sharpe from Qlib's information_ratio or mean / std (the current feature is a monotone function of ARR, which is already slot 4, and carries no volatility information), say so and I will adjust.
4. test: test/qlib/test_bandit_metrics.py feeds the extractor a Series with the exact key names Qlib logs and checks that every field is read, that a missing key produces a warning naming it, that a missing drawdown stays neutral, and that changing ARR moves EnvController.reward. The first test fails on the old key. All tests are marked offline so they run in CI.

Reward vector semantics are otherwise unchanged: same eight slots, same weights, same -mdd sign convention. log/ui/app.py still has its own copy of the key list; left alone here to keep the diff focused, happy to switch it to the constants in a follow-up.

Checks

• pytest test/qlib -q: 12 passed
• black --check -l 120 . and isort --check . as in the Makefile: clean

🤖 Generated with Claude Code

https://claude.ai/code/session_011Rf4jRK6VVz9qZQ9fcDErw

---

📚 Documentation preview 📚: https://RDAgent--1479.org.readthedocs.build/en/1479/

[michaelraziel-bit]

@microsoft-github-policy-service agree