fix(#934): defer plan authorship under --start-paused to the first tick

[0frankie]

Closes #934.

What

Under --brain llm --start-paused, boot built the default world and the LLMPlanner authored every default persona's day at attach time — real plan calls — then a pre-start POST /config threw that world away and re-authored for the applied cast. ~$0.13 (~2%) wasted on showcase candidate 1, scaling with default-cast size and the plan model's price, and the discarded personas muddied the lifetime by_actor ledger.

Fix

The issue's suggested shape: PennStepper grows defer_plans (wired to --start-paused in main()).

• attach_agents holds each agent on its safe static day (source "deferred" — the same fallback an unusable model answer gets). Seeding still runs its planner-present branches: the planner will read that memory, just later.
• The first tick() authors for the cast that is actually running — behind the budget guard, under the app lock — then re-stamps the manifest so the stored provenance describes the authored day. Every rebuild re-arms the deferral, so a pre-start re-cast stays free.
• The authoring block is extracted from attach_agents into cognition.author_llm_plan and shared by both paths, so they cannot drift.
• The mock path is unchanged (MockPlanner is free); the boot PLAN line says "deferred" instead of misreporting 0/N.

Verification

• Red: new test_paused_boot_plans_934.py first pinned today's waste (boot spends day_outline plan-role calls with defer_plans=True requested).
• Green: 4 tests — no plan spend until the first tick + authored-exactly-once, undeferred boot unchanged, manifest re-stamp (via a real RunStore), mock unaffected.
• Full suite 2472 passed / 2 skipped (test_godot_smoke deselected — it times out in any fresh worktree with no imported .godot cache, including on the untouched base commit).

🤖 Generated with Claude Code