[BUG FIX] [MER-5885] Restore UpGrade data capture parity for native A/B experiments

[eliknebel]

https://eliterate.atlassian.net/browse/MER-5885

Prerequisite work: This PR is based on PR #6786 and PR #6784. Both prerequisites have been squash-merged into hotfix-v0.34.1. This branch has been rebuilt on those squash commits, the superseded prerequisite commits were removed, and the complete MER-5885 diff was reverified.

Summary

Restores the experiment analytics data needed for UpGrade-style research and analysis while keeping raw learner outcomes independent from optional causal attribution.

Experiment evidence

• restores focused experiment attribution for evaluated part, activity, and page attempts
• restores attribution for media statements using authoritative attempt or published-section content
• preserves weighted-random assignment scopes and existing Thompson Sampling outcome, reward, and policy-update semantics
• carries complete experiment, assignment, intervention, condition, enrollment, project, publication, algorithm, policy-version, revision, and attempt identities where applicable
• fails safely when assignment or selected-branch evidence cannot be established, without suppressing the underlying xAPI statement
• keeps explicit experiment rewards separate from attempt scores; missing reward_value remains NULL, while explicit values including 0.0 are preserved

Initial condition assignments

• adds a dedicated experiment_condition_assigned xAPI statement and raw-event type
• emits exactly once when a new assignment is persisted and does not emit again for sticky reuse
• records the exact persisted assigned_at value and links the attribution to the dedicated statement hash
• supports weighted-random and Thompson Sampling assignments across intervention and section-enrollment scopes

Raw analytics and ClickHouse parity

• propagates pseudonymous enrollment_id across applicable attempt, page-view, tutor-message, and video statements
• preserves Torus account IDs in raw_events.user_id without falling back to actor email/mbox values
• adds the nullable raw-event and experiment-attribution fields required for analysis and historical replay
• normalizes projection behavior across direct upload, Lambda transformation, and replay/backfill
• renames the attribution relationship column to raw_event_type for clarity before production release
• consolidates the unreleased ClickHouse experiment migrations into the original migration, consistent with the planned QA database reset
• adds a deterministic, project-scoped ClickHouse query that reconstructs the UpGrade v0.33.0 compatibility dataset from raw outcomes and assignment evidence

Researcher and operator experience

• adds researcher documentation covering experiment workflow, terminology, S3 xAPI shape, ClickHouse fields, stable keys, event/attribution relationships, privacy boundaries, and example analysis queries
• displays the experiment UUID in experiment details and improves the responsive, accessible details layout
• constrains ClickHouse confirmation dialogs to the established modal sizing pattern

Important design boundaries

• PostgreSQL remains the runtime source of truth; S3 xAPI remains the replay source; ClickHouse remains an analytical projection
• raw evaluated outcomes are captured whether or not experiment attribution succeeds
• condition-assignment delivery remains best-effort; a transactional outbox, reconciliation, and historical assignment backfill are deferred
• exact attempt-time publication provenance remains the separately tracked MER-5889 limitation
• S3 xAPI may contain sensitive learner information; ClickHouse projections use bounded identifiers, including linkable pseudonymous enrollment IDs, and remain subject to approved analytics access controls

Verification

• 239 affected Elixir/LiveView tests passed, with 1 excluded
• 26 Lambda ETL tests passed
• mix compile
• mix format --check-formatted
• git diff --check
• requirements traceability passed for FDD, plan, and implementation evidence
• harness work-item validation passed
• security, performance, Elixir, UI, TypeScript, and requirements reviews completed with no remaining findings

Reconciliation status

• PR [BUG FIX] [MER-5884] Lambda ETL does not insert experiment attribution records into click house #6784 squash-merged as 1c14fc71c9
• PR [FEATURE] [MER-5883] Add configurable weighted random assignment scope #6786 squash-merged as 539b3c98d5
• Branch rebuilt on updated hotfix-v0.34.1
• Superseded unsquashed prerequisite commits removed
• git range-diff confirmed all retained MER-5885 commits were patch-equivalent
• Final-base verification and reviews rerun successfully

[github-actions]

	Warnings
⚠️	PR is large (6515 LOC changed). Consider splitting.

Risk score: 24 → risk/high

Generated by 🚫 dangerJS against 59c9003

[github-actions]

AI Review — elixir

No issues found

[github-actions]

AI Review — performance

No issues found

[github-actions]

AI Review — ui

No issues found

[github-actions]

AI Review — typescript

No issues found

[github-actions]

AI Review — security

No issues found

[nicocirio]

I think this query cannot currently reconstruct the compatibility dataset from the activity events we emit. It filters verb_id = '.../evaluated', but ActivityAttemptEvaluated.new/3 emits .../completed and both direct/Lambda project that as activity_attempt.

The producer test currently asserts the enrollment extension but not the emitted verb, while the compatibility-query test uses hand-built raw-event fixtures whose default verb is .../evaluated. Could we make these paths share one contract—align the query with the emitted verb and build the compatibility fixture from ActivityAttemptEvaluated.new/3 (or assert the same canonical verb in both tests)—so this cannot drift again?

[nicocirio]

If I understand correctly, a section may have active experiments at different Decision Points. This compatibility query only takes section_id as scope, unlike the query cookbook (docs/user/experiments/queries.md), which scopes experiment analysis by experiment_id. Could we either add an experiment_id parameter here or explicitly document that the query picks the most recent assignment for an enrollment regardless of experiment? Otherwise a researcher could silently attach a condition from a different concurrent experiment to an outcome.

[nicocirio]

hydrate_assignment_match/1 here is nearly identical to the one in media_attributions.ex, apart from section_slug. Would it make sense to extract the common hydration step so the two projections cannot drift the next time their selected fields change?

[nicocirio]

Small consistency note, not a bug: the single-assignment path receives {:ok, decision} for sticky reuse and therefore never calls its emission callback, while the batch path reaches ConditionAssignmentEmitter for the same case and relies on its reused?: true clause to return :ok. Both paths are covered and correct today. Would it be worth converging on one place to own that guard when this flow is next touched?