feat: tolerate recoverable seat intent inputs

src/cognitive_agent/infrastructure/llm/openai_responses.py

@@ -139,6 +139,11 @@ with this conservative deterministic baseline: longitudinal 80%, cushion tilt 30
 remaining upright and reversible. State this inferred ergonomic baseline in decision_summary. Apply
 the same causal reasoning to paraphrases such as insufficient legroom, feeling squeezed, wanting more
 space, or asking for a comfortable posture; do not require the user to name a motor or percentage.
+When a user simply asks to make the seat more comfortable without numeric targets, use that same
+80/30/35/10 ergonomic baseline. For an unsaved "boss", "executive", or equivalent front-seat
+clearance mode, infer the conventional intent of moving this controllable front passenger seat out
+of the way and use this deterministic baseline: longitudinal 20%, cushion tilt 30%, backrest recline
+10%, and legrest height 0%. State the inferred mode meaning and targets in decision_summary.
 
 When execution_policy.planning_environment is dry_run_candidate_synthesis, a named preset is not
 required. For a commonly understood mode, activity, or comfort goal, decompose the goal into the
@@ -156,8 +161,10 @@ capabilities as usual.
 
 Return every understood goal in requests using an exact capability name from the registry. Use
 target_scope and operation to preserve the user's intent. Put concrete capability parameters in the
-parameters list when they are known or defensibly inferred. Do not drop catalog_only requests; the
-runtime router, not the model, decides which domain adapter can execute them.
+parameters list when they are known or defensibly inferred. Preserve a catalog_only request when the
+user explicitly asks for that exact feature, but do not invent catalog_only capabilities while
+synthesizing a vague comfort goal or an unknown named mode. Inferred seat postures must use only the
+registered executable seat motion capabilities so the plan can safely reach hardware.
 For executable capabilities, resolve follow-up and relative intent against runtime state and return
 absolute target parameter values that the adapter can execute. Do not delegate that semantic decision
 to downstream code. Catalog-only requests may remain relative when their current state is unavailable.