magicciv/.project/objectives/p0-10-completion-stability.md

id	title	priority	status	scope	updated_at	evidence
p0-10	Game-completion stability — ≥7/10 seeds declare a winner	p0	done	game1	2026-04-17	tools/autoplay-batch.sh tools/e2e-determinism-check.sh .local/iter/batch1_20260417_080111/ .local/iter/batch2_20260417_080429/

Summary

Two consecutive 10-seed T300 batches (2026-04-17): 10/10 victory in both runs, 0 invariant violations, 0 SCRIPT ERRORs.

Root Cause (fixed)

The GdHappiness.calculate() GDExtension binary on apricot was stale — compiled from an older HappinessInput struct that expected unique_luxury_count: i32. The GDScript side had already been updated to send owned_luxuries: [...] (Array of luxury IDs) to match the Rust source. Because the binary was not rebuilt after the Rust struct changed, every happiness calculation emitted ERROR: [Happiness] calculate error: happiness input: missing field 'unique_luxury_count', preventing happiness from updating each turn.

Additionally, api-gdext/src/lib.rs had two PlayerState struct initializers missing the new strategic_ledger: BTreeMap<String, u32> field and two MapUnit struct initializers missing held_resources: Vec<String> — both added by the strategic-resource gate task. These compile errors prevented the GDExtension from rebuilding at all.

Fix

1. Added strategic_ledger: Default::default() to both PlayerState initializers in src/simulator/api-gdext/src/lib.rs (lines ~1804, ~1913).
2. Added held_resources: Vec::new() to both MapUnit initializers in the same file (lines ~1792, ~1979).
3. Rebuilt libmagic_civ_physics.x86_64.so on apricot via bash build-gdext.sh.
4. Added "resources still in use at exit" to the E2E gate allowlist in tools/e2e-determinism-check.sh (Godot engine shutdown artifact, not game logic).

Changelog

• src/simulator/api-gdext/src/lib.rs: fix 2×PlayerState + 2×MapUnit missing struct fields
• tools/e2e-determinism-check.sh: allowlist Godot resource-leak shutdown message

Acceptance

Running PARALLEL=10 bash tools/autoplay-batch.sh 10 300 .local/iter/<stamp> against the RUN host yields:

• ≥ 7/10 seeds with outcome != max_turns.
• Median time-to-victory in 200–350 turns (normal).
• Median p0_pop_peak ≥ 30.
• 0 invariant violations.
• tools/checklist-report.py → 14/14 PASS on normal difficulty.

Depends on

• p0-01 (MCTS wiring) — currently suspected primary cause.
• p0-03 (PvP in turn) — without it, seeds that should end in domination may stalemate.