Generated Compiler Output Staleness

[FHoffarth]

Explain, before fixing, why KnowledgeDBCompiler output diverges from the committed artifacts.

Scope

Two committed artifacts regenerate on every full test suite run and differ semantically from HEAD — this is genuine content divergence, not newline drift:

File	Delta
data/cruise_intelligence_db.json	+14,703 bytes
data/cruise_knowledge_graph.json	+15,583 bytes

They are the last known source of working-tree churn after a full run.

Classify the divergence semantic vs serialization, per top-level key, before proposing any refresh.

Non-goals

• No refresh commit in this issue
• No compiler refactor
• No json.dump migration — 28 direct call sites remain across src/, scripts/ and tools/, deliberately out of scope

Acceptance criteria

• Per-key classification of the diff (semantic vs serialization)
• A stated cause: stale commit, changed inputs, or non-determinism
• Confirmation of whether compiler output is deterministic across two independent runs
• A recommendation naming the exact files a refresh would touch

Evidence gates

• Diff produced against git show HEAD:<path>, not against a prior working tree
• Determinism checked by two independent runs into temp directories
• No in-place regeneration during the audit

Context

Repeatedly observed and deliberately excluded from PRs #29 and #30, where the two files were restored rather than committed. _load_ships() reads only identity/cabins/venues/operations/negative_intelligence.json, so the Meraviglia manifest and reports are confirmed not to be compiler inputs.

Label audit was approved for this issue but does not exist in the repository.