[Tristan-Stoltz-ERC]

Summary

Stacked on #2852. Adds a preregistered, simulation-first experiment to measure whether validated controller-outcome feedback improves recovery without weakening the controller authority boundary.

Added

• docs/research/ROBOTICS_PAIRED_RECOVERY_BENCHMARK_V0_1.md: four-arm design (native-controller reference, Symthaea intent, guarded adapter, guarded feedback loop); six fault families; non-compensatory lifecycle/dispatch gates; paired statistics; evidence tiers and claim limits.
• docs/research/robotics-paired-recovery-scenarios-v0.1.json: frozen 30 nominal seed cases and 6 × 3 × 30 = 540 fault cases, for 570 paired scenario identities and 2,280 arm-episode records across four arms. Numeric values/timing that depend on the controller profile must be frozen before scored runs.
• CI trigger allowlist expanded for the stacked outcome-ledger, paired benchmark, and scorer branches.
• Rebased onto the latest outcome-ledger head, so the PR stack no longer diverges from its base.

Research hypothesis (not a result)

On the same scenario/seed, B3 should improve fault-recovery success against B2 while keeping nominal success within a preregistered 1 percentage-point non-inferiority margin. No weighted reward can compensate for a hard boundary violation. A proposed +10 percentage-point recovery improvement is a target, not a result or industry standard.

Claim ceiling

This PR defines the experiment; it does not implement a rollout runner, restart-safe persistence, a physical robot adapter, or a safety certification. The current in-memory ledger does not by itself provide durable restart fencing. Any missing, crashed, timed-out, or unresolved episode must be represented, not discarded.

Validation

• Manifest JSON syntax, unique seed counts, 6×3×30 fault expansion, and 570/2,280 aggregate counts were checked author-time.
• Current exact PR head: 11937c3
• Exact-head CI run: https://github.com/Luminous-Dynamics/symtropy/actions/runs/40880783580 (queued, not passed).
• No simulation outcomes or performance gains are claimed in this PR.

Related: #2828, #2836, #2847, #2852.