[P3] release.yml export jobs upload artifacts without post-export existence checks

[itsmiso-ai]

Ask: Add post-export verification to the release workflow so a broken Godot export can't ship as a tagged release asset.

Expected files: .github/workflows/release.yml

Problem:
The release workflow (release.yml) runs the Godot export step and immediately uploads the resulting archive to the GitHub release. There is no post-export assertion that the produced archive actually contains the artifacts Godot was supposed to write. The test workflow (test.yml) already has a precedent: its export-validation-* jobs explicitly check for the binary/PCK/wasm/html files before declaring success.

The release workflow's validate job runs the test suite but does not re-verify the platform exports — a corrupt or empty export that still returns exit code 0 from --export-release would be uploaded and published as a release asset.

Evidence:

• .github/workflows/release.yml — each Export ... build step is followed directly by a Python zip step and softprops/action-gh-release upload; no file-existence or non-empty checks before upload.
• .github/workflows/test.yml — export-validation-linux/windows/macos/web jobs each have a "Verify export output" step that asserts the artifact exists and is non-empty; the release workflow has no equivalent.

Acceptance:

• After each Export ... build step in release.yml, an assertion step confirms the expected files exist and are non-empty before the archive is built and uploaded.
• A failed export fails the release workflow (matching the existing set -euo pipefail posture).
• A release that previously would have shipped a missing or empty asset now fails fast.

[itsmiso-ai]

Not worth doing yet

Considered during this audit but not actionable right now — listed here so the next runner has the context:

• Save debounce window widening: the PERSIST_INTERVAL_TICKS = 10 debounce in scripts/main.gd (persist) can lose up to ~9 ticks of progress on a hard crash. Acceptable today because the loop is desktop-resident and the player can manually save / quit to flush. Revisit only if on_quit / crash durability becomes a complaint.
• Goal catalog / reward catalog drift risk: RotatingGoal.GOAL_CATALOG and GoalReward.REWARD_CATALOG are coupled by string ID. A unified catalog is appealing but the two files are intentionally split (RotatingGoal has no reward knowledge, GoalReward has no goal knowledge). Consolidating would create a new circular dependency for a coupling risk that has not bitten yet.
• Manual release tag format parity with release-please: the regex in manual-release.yml already accepts plain semver with optional pre-release/build metadata, which matches what release-please produces. No drift to fix today.
• Worker cap / active worker count edge cases: WorkerCapLogic and MilestoneManager agree on the active-worker definition, and existing tests cover the boundary. No new finding beyond what closed audits already captured.
• Render overlay scaling for very small tile sizes: worker_collision_offset floors spacing at 8px; below ~20px tile size the collision slots overlap visibly. Not a real-world issue at the pinned 48.5–58px tile range. Revisit only if zoom-out (≤ 0.5x) ships.
• PR-reviewer-action gate for approve verdicts: previously raised in [P3] AI PR review check fails on approve verdicts — bot token lacks create/approve-PR permission #348 and intentionally deferred (needs-human). Not re-filed.

[itsmiso-ai]

Grooming pass: this issue is well-specified and ready. The ask (add Verify export output steps to each export job in release.yml, modeled on the existing ones in test.yml) is determinate and scoped to a single file. The two implementation notes in the body (zip heredoc placement, macOS conditional pck write) are the right gotchas to flag. No blockers, no missing info — good to pick up.