Fix random problem selection by difficulty

[Ericko00o]

Explicit random selection currently shares the spaced-review recommendation policy. When multiple problems are overdue, repeated clicks can cycle through those problems indefinitely, excluding unseen problems and selecting reviews outside the chosen difficulties.

Separate explicit random selection from automatic recommendations. Random draws uniformly sample the problems matching the selected difficulties and topic filter, excluding the current problem whenever an alternative exists. Initial recommendations retain overdue-review priority across the broader review pool. Random selections also display the problem's difficulty.

Add regression tests for unseen-problem eligibility, difficulty filtering, current-problem exclusion, single-problem and empty pools, and the distinction between filtered practice and broader review recommendations. Update the lobby's expected selection text and add browser coverage for random selection, page restoration, and difficulty changes.

Validation: node --test tests/browser/problem-picker.test.js tests/browser/lobby.test.js completed with 43 tests passing, no failures, and 112 browser tests skipped because Playwright Chromium was unavailable. The JavaScript syntax check and git diff --cached --check passed. The required ./scripts/test.sh gate has not been run, and browser interaction remains unverified locally.

Closes #210

---

Summary by cubic

Separates explicit random selection from the spaced-review recommendation policy so random draws sample uniformly from the problems matching the selected difficulties and topic filter, show the drawn problem's difficulty, and exclude the current problem whenever an alternative exists. Recommendations keep overdue-review priority across the broader review pool.

• A random draw persists across page restores and history maintenance such as deleting reports or merging older account history, and across returning from an unfinished interview; explicit difficulty and duration choices are also restored when the lobby is rebuilt. Only a newly graded attempt at the drawn problem ends the draw and returns to recommendations.
• A tab-local draw identity makes completion exact even under a backward clock adjustment, and blocked storage leaves navigation and report saving usable.
• A difficulty change before history finishes loading no longer resets the picker to recommendations.
• Adds regression tests for unseen-problem eligibility, difficulty filtering, exclusion, narrowed-practice pools, empty pools, topic resets, real Back navigation, and difficulty-change, history-change, page-restore, interview-navigation, draw-completion, and storage-failure behavior.

^{Written for commit d6f4ca2. Summary will update on new commits.}

[jserv]

pickerMode is set to "random" and never goes back to "recommend", so after one click every later recommend() draws uniformly: filterSelectionChanged(), sign-in, report deletion and every settle(). Concrete case: click Random and get X, finish X, press Back. The bfcache restore refetches reports that now record X as passed, yet roll, avoidedProblem and the pool are unchanged, so settle() offers X again (and silently drops the reason keptDraw is cleared there). Reset to "recommend" wherever avoidedProblem is reset to undefined, and in settle() whenever the reports differ from the ones the random draw was made from (track them the way keptDraw does).

[jserv]

Check CONTRIBUTING.md carefully and enforce the rules. In particular, the manner for writing git commit messages.

[jserv]

Before asking someone to review, you MUST resolve any regressions flagged by GitHub Actions.

[ColtenOuO]

A back/forward restore does not keep the draw: the pageshow handler clears problemTopic but leaves pickerMode, roll and avoidedProblem as they were. After Random under a topic filter, going to preflight and back re-runs the same roll over the unfiltered pool and selects a different problem. Either keeping the topic on restore, or resetting the draw there the way the topic-change handler does, would make the two agree.

[jserv]

Fold similar commits and refine git commit messages per CONTRIBUTING.md.

[jserv]

Rebase latest main branch and resolve conflicts.