PokeSwift Extractor Contract Change

Read AGENTS.md and SWIFT_PORT.md before changing milestone-sensitive behavior.

Treat generated content as evidence, not authoring surface. Prefer fixing the first broken contract layer instead of compensating in runtime code.

Workflow

1. Define the failing contract. Capture the exact value or shape that is wrong: missing dialogue ID, wrong warp/checkpoint, broken map script, bad trainer manifest, missing audio cue metadata, stale item/species/move field, or loader decode mismatch.

2. Trace the first diverging layer. Use this order: pret/pokered asm/data -> PokeExtractCLI -> Content/Red -> PokeDataModel/PokeContent -> PokeCore/PokeUI. Fix the first layer that differs from the expected source-driven result.

3. Pick the right ownership boundary. If the extracted value is wrong, change PokeExtractCLI. If the extracted value is right but the shared model cannot represent it, change PokeDataModel first. If the model is right but loading or validation is wrong, change PokeContent. Only change PokeCore or PokeUI after the upstream contract is correct.

4. Regenerate instead of patching artifacts. After extractor changes, run ./scripts/extract_red.sh. Never hand-edit Content/Red/** as the fix.

5. Validate upstream first, then downstream. Run focused PokeExtractCLITests and PokeContentTests. Then run the smallest runtime or UI tests that consume the changed contract. Update SWIFT_PORT.md if the shipped scope, milestone boundary, or system ledger changed.

Ownership Guide

• PokeExtractCLI Own parsing, normalization, template assembly, and deterministic manifest generation. Search here for battle text templates, trainer dialogues, encounter tables, map objects/scripts, warps, marts, items, audio cues, and source-derived flags.

• Content/Red Treat as generated output. Diff it to confirm what changed, but do not author fixes here.

• PokeDataModel Own shared manifests and save/telemetry schemas. Put contract changes here before duplicating structure in loaders or runtime code.

• PokeContent Own content loading, validation, and repo-contract tests. Fix this layer when manifests are correct but decoding, validation, or loader defaults are wrong.

• PokeCore and PokeUI Consume contracts. Change these only after upstream data and shared types are proven correct.

Common Triggers

• wrong Pokecenter blackout checkpoint or player start location
• missing trainer lose dialogue or battle text token
• battle text template drift against common_text.asm
• encounter table, mart, item, species, move, or evolution manifest mismatch
• audio metadata or dialogue-event extraction mismatch
• schema bumps that require loader and save defaults
• Content/Red diffs that look suspicious after extractor work

File Checklist

• Read Sources/PokeExtractCLI for parser and manifest generation ownership.
• Read Sources/PokeDataModel for shared schemas.
• Read Sources/PokeContent for loader and contract validation ownership.
• Inspect Content/Red only to confirm generated output, never as the source of truth.
• Read SWIFT_PORT.md if the contract change alters shipped scope or milestone claims.

Validation Matrix

• If extraction logic changed: Run ./scripts/extract_red.sh.

• If manifests or schemas changed: Run focused PokeExtractCLITests and PokeContentTests.

• If runtime consumers changed because of the contract: Run the smallest relevant PokeCoreTests or PokeUITests slice.

• If file moves or target-sensitive edits happened during the change: Run tuist generate --no-open before broader validation.

Failure Shields

• Do not hand-edit Content/Red/**.
• Do not bury schema drift in runtime if statements.
• Do not change runtime behavior first when the extractor output is already wrong.
• Do not stop at compile success; prove regenerated artifacts and loaders both validate.
• Do not forget save/schema coupling when shared models change.

Review Mode

When reviewing an extractor or contract patch:

• Start from git diff --stat and the changed generated artifacts.
• Check whether the extractor output matches the expected Red source or just "looks plausible."
• Look for missing shared-model updates when new fields appear.
• Look for loader defaults or decode paths that silently discard new data.
• Treat unexplained Content/Red diffs as first-class findings.

Output Expectations

Report these points clearly:

• expected source-driven contract
• first diverging layer
• files changed and why
• regenerated artifacts touched
• upstream and downstream validation run
• whether SWIFT_PORT.md changed