Pose each problem as an interview scenario

[jserv]

A candidate shown a practice problem as it is published recognises it and recites a remembered answer, which is not what an interview measures. This branch poses every problem as a workplace scenario written around the same judged contract: a renamed entry point or class, and renamed parameters where the published ones identify it, a brief that leaves the limits to be asked for, clarifications and follow-ups the interviewer can use, a private contract, and a three-rung hint ladder that log_hint serves one rung at a time, withholding the last until the candidate has shown an approach. The interview page names the published title once in small print, so the problem can be found again afterwards, but never its number, statement or examples; the interviewer is told never to name it. The browser knows a problem by its page name, and the published id lives in one map that resolves old links and old history and backs the lobby's opt-in toggle. Solution notes reach only the report prompt, which may not name the published problem back to the candidate. The contract moves to bundle 5, and bundle 4 stays scorable because the rubric and schema did not change.

The generator holds the disguise: it refuses a variant whose text, starters, constraints, parameter names or case labels still name the source, a page example that is a published example or reuses one of its arguments, and a hint or clarification that walks through one. Where a judge had no unpublished case worth showing, it gained one, each checked against solutions written blind from the briefs.

The REACTO checklist is ticked from the interviewer's recorded evidence, and a session once ticked Coding with nothing typed. The server now refuses Coding, Test and Optimizations evidence until the editor holds code the candidate added beyond the starter, counted in order so deleting the starter's comment neither counts nor cancels a one-line answer. The browser sends each language's starter beside the buffer, so a tab switch, a tab round trip or a paste into a cleared editor is measured correctly.

Four smaller changes come first, each standing on its own: a test setup error is told apart from a failing test, passed problems return for spaced review, a candidate can record an optional readiness check-in that never reaches the interviewer or the score, and a report's recurring improvement item can be shared as a practice focus that shapes at most one neutral follow-up.

Verified with ./scripts/test.sh at the tip: exit 0, 481 browser tests and none skipped. The series was rebuilt with the full gate at the scenario and REACTO commits, and the drift checks, generator tests, Rust web and agent tests and the page-shape browser suites at each commit between them. Every variant was solved from its brief alone against its judge, 150 of 150, and every judge case added since was run against those solutions. scripts/interview-behavior-check.sh passes against live Gemini with the current prompts. CI's check and mutants lanes pass.

Not run: the credentialed voice and parity checks, which need a live LiveKit room, and scripts/visual-parity-check.sh, whose golden images were already stale on main.

---

Summary by cubic

Poses each practice problem as a workplace scenario around the same judged contract, so a candidate who recognizes the source problem can't recite a remembered answer. Bundle 5 replaces bundle 4; bundle 4 stays scorable.

Main change

• Each scenario adds a renamed entry point or class, a brief, clarifications, and follow-ups returned when the coding round completes; a restarted interviewer after Test and Optimizations is told the problem is solved, not asked to return to coding. A three-rung hint ladder is served one rung per request, holding the last until the candidate shows an approach; a request with no clue isn't counted.
• The browser and interview know only the scenario's page name; one map resolves old links and rewrites local history once. Account reports are stored under the stable published id and read back under today's page name, so a later scenario rename keeps old reports working. The page shows the published title once in small print, and the lobby can reveal the source on request.
• The generator refuses variants that name the source in any candidate-visible text or starter, and blocks C starters that omit the returned-array malloc note.
• Pages never show a published example; they show unpublished judge cases, adding new ones when the bank has none.
• Solution notes reach only the report prompt, which never points the candidate back to the published problem.
• Coding, Test, and Optimizations evidence requires code typed beyond the starter; skips still count.

Smaller changes

• A test setup error is told apart from a failing test.
• Passed problems return after 1, 3, 7, 14, then 30 days, and the lobby says when review is due.
• A recurring report improvement can be shared as a practice focus that shapes at most one neutral follow-up, passed via session storage.
• The generator is split into a problem_bank package and each rule step is tested at its threshold.

^{Written for commit 3027bb8. Summary will update on new commits.}