Connect reading, saved vocabulary, review and conversation into one learning loop

## Priority

P2 — focused reliability or product improvement

## Context

Portfolio audit follow-up (2026-09-08). The next useful product milestone is one language/level working end to end. Existing feature-level issues are mostly closed; the gap to validate is the transition between features and durable progress.

## Starting points / evidence

Depends on the storage migration and provider fixes. Coordinate with existing conversation history #20 and real SRS integration #33, both closed, as implementation history.

## Implementation scope

Choose a currently supported language/level for a fixture-driven vertical slice: read content, save a word with its context, schedule/review it, then use it in a conversation. Keep AI feedback structured and reviewed against a small checked content bank. Build on existing SRS/history rather than a parallel storage system. Prepare return-use/vocabulary-retention pilot measures without claiming learning efficacy.

## Acceptance criteria

- [ ] A complete fixture flow preserves language, word identity, source context and SRS state across reloads.
- [ ] Conversation practice can use selected learned words without inventing new progress records.
- [ ] Review scheduling and grading behavior have deterministic tests with mocked model output.
- [ ] User-facing empty/error states explain the next action and allow retry without duplicate words.
- [ ] Document a consented pilot protocol and success criteria before subscription work.

Keep the change focused on this issue. Use synthetic fixtures for sensitive data and mocked providers for automated tests; do not require paid model calls to run the regression suite.

Start by reviewing the storage migration and the implementation history in closed issues #20 and #33. Trace one supported language/level through reading, vocabulary saving, SRS review, and conversation using synthetic fixtures and mocked providers. Done means the end-to-end state survives reloads, tests are deterministic, retry and empty states are covered, and the consented pilot protocol is documented.

Written by the indexing model from the issue text.