Admin: Add arbitrary dataset-revision comparison and field-level diff views

[gekichumai-admin]

Phase / step: Phase 4 — Provenance / Step 3
Parent: #301
Depends on: #335, #330

Context and goal

Dataset publications form an immutable revision history. Administrators must be able to compare any two retained revisions for a chart—not only adjacent revisions—and understand the data/evidence changes that produced each field change.

Scope and implementation notes

• Add a dataset-publication timeline for a chart with immutable revision identity, publication timestamp, completeness marker, and a concise change summary.
• Let the user select arbitrary base and comparison revisions, swap them, and copy a URL that restores both selections.
• Render field-level added, removed, and changed values using canonical paths. Include resolution-state, chosen-source, candidate, rule, and reason changes when those differ even if the final value is unchanged.
• Provide summary counts and filters for value changes, evidence-only changes, provenance gaps, and manual overrides.
• Use appropriate comparison presentation by value shape: concise inline diff for scalars and structured before/after views for arrays/objects. Do not reduce structured values to unreadable JSON blobs by default.
• Link each changed side to its field provenance drawer and retained raw evidence.
• Mark best-effort backfilled revisions and missing historical evidence explicitly. Never invent a source or silently treat missing history as unchanged.
• Handle chart creation/removal between selected revisions and revisions where the chart identity cannot be resolved.

Acceptance criteria

• Any two retained dataset-publication revisions can be selected and compared in either direction.
• Revision selections and active diff filters survive reload through URL parameters.
• Added, removed, value-changed, and evidence-only-changed fields are represented correctly.
• Candidate, selection-rule, decision-reason, and source changes are visible even when the chosen value is equal.
• Each diff row links to the corresponding provenance detail and raw evidence when available.
• Incomplete historical data is labelled rather than inferred.
• Creation, removal, and unresolved identity cases have dedicated states.
• Large diffs remain navigable with filtering and incremental rendering/pagination as supported by the API.
• Component tests cover scalar, nested, array, evidence-only, unknown, and historical-incomplete changes.

Out of scope

• Restoring an old revision as the active dataset.
• Comparing entire datasets across all charts in one browser view.
• Editing revision metadata.
• Data-publication controls.