[P0][Plan 1] SF-002 — Mark review stale when generation inputs change

[Ankit6149]

Confirmed problem

After a campaign is generated, changing notes, audience, links, repository, uploaded source text, selected destinations, provider, model, or endpoint does not invalidate the existing review. Users can unknowingly review, export, or publish drafts produced from an older source.

Scope

• Define a normalized source snapshot and stable fingerprint.
• Store the fingerprint on every generation run.
• Recompute it whenever a generation-relevant field changes.
• Keep old output visible but mark it as generated from an older snapshot.
• Block final export and publishing while stale unless an explicitly designed archival action is used.

Acceptance criteria

• Source A is generated, then changing to Source B immediately marks the campaign stale.
• A persistent warning remains visible in Review; a temporary toast is insufficient.
• Stale drafts cannot be published or exported as current work.
• Regenerating from the current source clears the stale state.
• The comparison is deterministic across save/reopen.

Required verification

• Reducer/domain tests for every generation-relevant field.
• Save/reopen regression test.
• UI integration test for stale banner and blocked actions.
• Frontend tests and production build must pass.

Dependency

Start after the minimal canonical CampaignState, SourceSnapshot, and GenerationRun contracts exist.

[Ankit6149]

Completed in #54 and merged to master as f64485499ed0d5a20e47939ed45f46b90dddc510.

Verification:

• deterministic normalized source fingerprint;
• generation-run snapshot persistence and legacy restore;
• persistent accessible stale warning;
• stale copy/export/publish guards at both UI and handler boundaries;
• field-level, save/reopen, legacy, tamper, and UI integration regressions;
• frontend tests, dependency audit, production build, MCP tests, and existing Python tests passed in CI run 201;
• final Vercel deployment succeeded.

Issue #17 was deliberately left separate because its atomic response-validation scope is larger and was not required to claim this issue complete.