Tune narration prefetch endpoint with live-call budget.

Shipped as a39a589.

What Changed

This pass makes NPC narration more usable in the browser by adding cache-aware behavior, cache warming, and bounded live calls so the page remains responsive.

• Worked in job board and contract flow.
• Worked in Letta narration.

Why It Matters

LLM-backed flavor is valuable only if it does not stall play. The cache and prefetch work keeps the experience responsive while still supporting richer NPC beats.

Implementation Notes

The commit message for this slice was:

Prefetch now accepts max_live (default 1), reports deferred beats explicitly, and play-page warm-cache action uses the bounded mode by default.

Files touched in this slice included:

• public/api/narrate-prefetch.php
• public/play.php

Verification

• The change was committed to main and included in the running devlog timeline.
• The current archive test suite verifies devlog posts render with substantive body content, illustrations, screenshots, and valid asset references.
• For browser-facing slices, Playwright smoke checks in this build cycle covered the active game surfaces and devlog pages.

Visuals