create-kr-patch

Create a Korean patch for a retro game

Purpose

Base decisions on evidence from the target game. Keep the evidence and adopted results from the initial survey, fonts and encoding, PoC, extraction, translation, reinsertion, builds, distributable patches, and runtime verification traceable through completion.

Choose tools, languages, libraries, and fonts based on the target project's existing structure and the current environment. Regardless of those choices, preserve the required capabilities and verification criteria. Look up basic specifications from current primary sources when one straightforward search can recover them.

Treat source-language frequencies and structures from other titles as hypotheses until the target revision confirms them.

Foundational principle

Treat each playthrough as the whole product. Successful paths do not offset a known in-scope defect that can block progress, corrupt text or intended meaning, mislead the player about a choice, break established terminology or voice, or impair interaction. Fix it, present evidence and feasible options for an explicit human decision to narrow the scope or claim, or withhold the release. See references/strategy/build-and-verify.md §6 and references/strategy/translation-workflow.md §6.

Operating principles

1. Credit progress only to the claim it strengthens. Effort, plausibility, implementation cost, and local success do not strengthen another claim. Keep the hardest unresolved completion condition—or the one most likely to force redesign—visible. Choose work and verification scope from that condition rather than the size or order of edits; work elsewhere only to supply a prerequisite, advance another declared in-scope claim, or apply a recorded human decision. Among approaches that preserve the same prerequisites and proof scope, prefer the lower total cost of obtaining the evidence. See references/conventions/project-records.md §1 and references/strategy/build-and-verify.md §1.
2. Reassess when evidence conflicts or work stops producing information. An observation that contradicts an adopted explanation or completion claim, repeated equivalent evidence, local checks that do not strengthen the cumulative claim, or unexplained exceptions and conflicting accounts of current state are signals to stop. Reconstruct the affected baseline, completion condition, and live and rejected explanations before another attempt. Search references/tips/README.md by observation or affected transition when a close match could reduce investigation cost; after these signals, search before adding another local exception. Cases suggest hypotheses, not repairs; preserve valid evidence without forcing an analogy. See references/conventions/project-records.md §2 and references/strategy/debugging.md §2·§3·§6.
3. Keep product priorities and decision authority separate from implementation. Current implementation establishes behavior and change cost, not product value. The agent establishes technical facts and feasible options; humans choose product direction, quality, scope, acceptable loss, support, and investment. Routine implementation that preserves adopted intent remains the agent's responsibility. See references/conventions/project-records.md §1.1 and references/strategy/translation-workflow.md §4.1·§5.