✨ Add intermediate rewards with paper semantics

[flowerthrower]

🤖 AI text below 🤖

Description

Combines the intermediate-reward feature and the paper reward semantics in one top-of-stack PR while retaining the existing PredictorEnv lifecycle.

The public reward controls are intermediate_reward=True, reward_scale=1.0, and no_effect_penalty=-0.001. Setting intermediate_reward=False retains terminal-only rewards.

• comparable non-terminal expected-fidelity and ESP steps use exact or approximate figure-of-merit deltas;
• approximate expected fidelity uses translated target-basis gate counts and averaged calibration errors, including measurement;
• approximate ESP additionally applies the prototype's duration, parallelism, liveness, and coherence idle penalty;
• terminal ESP uses ASAP scheduling and stops charging a qubit after its final measurement or reset;
• critical-depth actions use exact non-terminal deltas, while estimated Hellinger distance remains neutral until termination;
• structural compilation progress and structural no-effect steps remain neutral;
• optimization no-effect and failed actions receive no_effect_penalty, with failures truncating the episode.

This is position 11 and the top of the stack. It depends directly on the GNN PR #788 and absorbs the former standalone intermediate-reward PR #760. No action, GNN, dependency, or experiment changes are included.

Validation

• repository lint;
• direct bit-exact comparisons with add-individual-bqskit-passes for approximate expected fidelity, approximate ESP, calibration maps, and terminal ESP;
• focused reward and timeout checks (21 passed, 25 deselected).

Fixes #670

Part of #664

Checklist

• The pull request only contains focused changes required for intermediate rewards and paper reward parity.
• Existing focused tests and release metadata cover the changed public behavior.
• I reviewed the complete base-to-head diff.

If PR contains AI-assisted content:

• Any agent that created, edited, or submitted GitHub content was explicitly authorized for that scope.
• Every agent-authored or agent-edited public text body begins with the visible disclosure 🤖 *AI text below* 🤖.
• I confirm that I have personally reviewed and understood all AI-generated content, and accept full responsibility for it.

[flowerthrower]

🤖 AI text below 🤖\n\nSuperseded by the rebuilt combined-reward layer at position 8.

[codecov]

Codecov Report

❌ Patch coverage is 38.37838% with 114 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/mqt/predictor/rl/approx_reward.py	13.8%	81 Missing ⚠️
src/mqt/predictor/rl/predictorenv.py	28.2%	33 Missing ⚠️

📢 Thoughts on this report? Let us know!