Spend and context visible to the model; one notice before auto-compaction

[Hmbown]

Gap.

• With auto-compaction on (the default), the context-pressure warning is suppressed (core/engine/compaction.rs:~150-172), so the model gets no notice before its memory is compacted.
• The model never sees spend so far, cost, or the remaining session budget. A token budget exists only while a goal is active (engine.rs:~3705-3728).

Change

1. Extend <turn_meta> (engine.rs:~3850-3887) with context: <pct>% of <window> and spent: <cost> this session. When the person has set a session budget, also add budget: <remaining> of <total>. Use values the usage and cost ledger already computes. Omit cost when pricing is unknown; never guess it.
2. Warn before compaction, even with auto-compaction on. Add a one-time notice per threshold crossing before "Making room" (the §19 wording): "Context is at 85%; Codewhale will make room soon. Checkpoint anything you need in files or the task list." After compaction the existing summary still says how to re-read live state. This is related to feat(engine): soft model-facing context notice while auto-compaction is enabled (no competing save/compact ceremony) #6425.
3. Optional session budget. Add a session_budget config key and a /budget command. On reaching it, stop the turn with a clear notice; never cut mid-tool silently.

Tests

• turn_meta_includes_context_and_spend
• unknown_pricing_omits_cost
• pre_compaction_notice_fires_once_with_auto_compact_on
• session_budget_stops_turn_with_notice

Size: M.