Classify context-specific translation adaptations

[jeprecated]

Note

This PR is stacked on top of #743. Until that migration PR is merged, GitHub will also show its commits in this PR.

See only the last two commits for the real changes. https://github.com/anki-geo/ultimate-geography/pull/744/changes/5e96018f74834bd751cfb67b9f1ece9fb3abdbd8..8bbeb7344ad6fc799f60f75fb98012f000a59590 (maaan I'm looking forward to Github getting Stacked diff support already!)

I wanted to separate the question of what kind of translation entry something is from the much riskier question of whether a translation is actually correct.

This PR therefore makes no learner-visible translation changes. It has two commits, both of which preserve the parsed CrowdAnki output for all 100 targets.

Commit 1: one simple context-specific adaptation

Taiwan is the small example that demonstrates the principle.

The existing Simplified Chinese description is not a reusable translation of the English sentence in every context. It is wording owned by that specific target-language field, with a geopolitical reason for being different.

The first commit moves it from the broad direct dictionary into a path-owned target_adaptations entry. That makes the difference, its expected English source, and its ownership explicit without changing the exported value.

Commit 2: apply the same principle across the deck

I then used AI to audit all 6,985 remaining translations.direct entries and apply the same distinction consistently across the language overlays.

The result moves:

• 116 direct source keys into 120 path-specific typed adaptations;
• one Hebrew field-specific value into a contextual translation.

Some source keys were used at more than one note path, which is why 116 keys become 120 explicit adaptations.

I was deliberately conservative about this. The audit retained 6,727 entries as direct translations, left 51 entries requiring native-language judgment alone, and left 17 possible canonical-source decisions alone. The 73 possible content errors found during that work are also excluded from this PR and handled separately in a draft follow-up.

The point here is not that AI has proven every translation semantically correct. The point is that it applied one explicit authoring rule across the existing data, and I can prove that this reclassification does not change what learners receive.

Verification

• Translation-profile consistency passes.
• Strict verification passes for all 74 main and 26 companion targets.
• All 100 targets were exported and compared as parsed CrowdAnki JSON.
• The aggregate output before and after both commits is identical.
• No note fields, GUIDs, deck/model identities, templates, tags, configuration, or media change.

This gives us a cleaner distinction between reusable translations, context-specific translations, and target-owned adaptations without combining that structural cleanup with editorial translation changes.