feat(calendar): grade natural-language user preferences

[Luiz do Valle (luizvalle)]

Replace the numeric (time slot, score) preference model and its ZOPA-based outcome_optimality metric with free-text preferences graded by per-task programmatic verifiers. The goal is to measure whether a model under test can follow preferences stated in natural language, and to surface where it fails.

A task now points at a preference.md written in the third person ("User prefers...", "User never..."). The loader inlines that text into the task, the assistant agent injects it into its user turn inside <user_preference> tags, and the system prompt explains the convention: never/always/only are hard constraints, prefers/would rather are soft, and an impossible request should be declined rather than forced.

Grading moves to evaluation/preference_adherence/. Each task registers a verifier that declares its hard predicates and weighted soft preferences; score_task enumerates the slots free on both calendars, filters them by the hard predicates, and reports:

hard_constraints_satisfied (0/1) - outcome respects every hard constraint,
including declining when no slot is feasible
soft_constraints_score (0-1) - chosen slot weight over best achievable

Scoring against the best achievable slot is what makes conflicting soft preferences well defined: a model is never penalised for a trade-off the calendar made unavoidable.

TimeSlotPreference stays on the models but is now inert - never prompted, never graded - because srbench-data-gen and the shipped small/medium/large datasets still generate it. Due diligence dispatches on preference type, since the reasonable-agent replay structurally requires numeric preferences.

Adds data/calendar-scheduling/soft_pref_demo: 12 tasks over a hardness spectrum, from a single soft preference to contradictory rules that can only be answered by declining.

Verified end-to-end against a locally hosted Qwen3.5-27B: 66.7% hard constraints satisfied, 0.667 mean soft score, and 0/2 impossible requests declined. The failures are informative - the model never declines, occasionally abandons a solvable task, and sometimes double-books the counterparty's calendar, which the existing has_conflicts check does not catch.

---

PR Checklist (do not remove):

• I linked this PR to an issue
• I included code instructions on how to test