feat(evaluation): report the reasoning-branch intervention delta as an eval metric

[gcordova10]

Problem

reasoning_intervention_delta (added in #109) measures whether the planner actually uses the reasoning
branch: bypass the branch, re-run the same batch, and report how far the trajectory moves. This is the
gate g discussed in #123. But nothing calls it during evaluation, so no checkpoint reports it —
today g is a manual, one-off measurement. #123 asked for the opposite: a number computed per
checkpoint that answers "is the reasoning used?" on its own, and rises by itself if the coupling ever
opens with data scale.

One caveat shaped the wiring: the function's forward did not pass the navigation inputs (#161) or a
fixed initial_noise. So a plain call would measure the delta off the real operating point — and, for
a stochastic planner, with different noise in the two runs, which swamps the signal.

Fix

• Extend the harness. reasoning_intervention_delta / horizon_intervention_delta take
  **forward_kwargs, threaded identically into both runs — a fixed initial_noise (so both runs share
  it) and the navigation inputs (so the delta is at the real operating point). Additive and
  backward-compatible.

• Wire it into eval. _evaluate_open_loop gains a report_intervention flag (default off).
  evaluate_il_policy turns it on; the training-time validation leaves it off, so training cost is
  unchanged. When on and the model has a reasoning head, it computes the delta on up to 50 batches
  (reusing that batch's stable initial_noise), averages, and returns reasoning_intervention_delta;
  _run_evaluation logs eval/reasoning_intervention_delta next to ADE/FDE.

• Guarded. With no reasoning head the metric is skipped, so this is a byte-identical no-op for
  non-reasoning checkpoints.

Verification

• ruff clean; mypy clean on Model/evaluation/faithfulness.py.
• Model/tests/test_faithfulness_forward_kwargs.py (new, 3 tests): a plain call is noise-dominated; a
  fixed initial_noise recovers only the intervention; the navigation inputs are threaded.
  test_reasoning_faithfulness.py stays green — 11 passed together.
• Model/tests/test_workflow_training_lifecycle.py (3 new tests): the metric is reported when opted in,
  absent without a reasoning head, and absent by default. These reuse the file's existing
  flytekit-gated eval harness, so they run in CI alongside the other workflow tests.

Coordination

As I flagged in #123, this touches _evaluate_open_loop, which #176 (feat/rollout-aligned-checkpoint-selection) also restructures. I offered there to base it on that branch or open against main and rebase once it lands — opening against main here so it is not blocked, and I will rebase onto #176 when it merges so the metric reports inside the composite selector. Happy to coordinate on the exact wiring point.

Refs #123 (the gate this computes) and #109 (the harness it calls).

[riita10069]

#168 (comment)

[gcordova10]

Rebased onto main now that #176 has landed, and MERGEABLE again. Both conflicts were additive rather than competing: each side added a keyword to _evaluate_open_loop, and #176 replaced the ADE/FDE computation with the 3-second window plus per-horizon accumulation, which this metric sits alongside rather than duplicates. ruff and mypy clean, 920 passed.

The description said I would rebase "so the metric reports inside the composite selector". The rebase makes the two coexist, but it does not put the delta into the selector: the composite score is computed in the per-epoch validation, where this metric is deliberately off. Wiring it there would add a per-epoch cost the current design avoids, so it reports next to ADE/FDE in _run_evaluation instead. Happy to wire it either way if you prefer.

On the scoping point: this does not add a reasoning layer or change what is trained — it is skipped when the model has no reasoning head, and the delta is capped at 50 samples, so Reactive-only runs take the path they take today. If you would rather park it until the reasoning layer is picked back up, that is fine by me; the rebase is done either way.