Add server-selected strict interview mode

[ZuoTso]

Summary

Introduce an opt-in Strict Interview mode for more honest self-assessment, reducing the ability to repeatedly practice a familiar problem through URL selection or page refresh.

Strict assigns problems on the server based on the selected difficulty, while Practice remains unchanged. A normal same-tab refresh starts a new attempt and excludes the previous assignment when its receipt is available. Session recovery is not included.

Testing

./scripts/test.sh passed (1,059 tests, no failures or skips; 46 hook checks). Manually verified with make web.

Closes #185

---

Summary by cubic

Adds an opt-in Strict Interview mode that assigns a matching problem server-side after media preflight, closing #185 by preventing repeated practice of a known problem via URL or refresh. Practice mode keeps its current behavior.

Selection and refresh behavior

• Difficulty remains a multi-select choice; the server draws from matching problems, excluding the previous assignment.
• A signed receipt in session storage remembers the previous problem per tab; clearing or replaying it resets the exclusion.
• Reloading returns to preflight, then starts a fresh attempt with a different problem; admission is refused with HTTP 409 if none remains.
• Strict ignores URL problem selection, does not fall back to offline practice, and restores no interview progress.
• Strict always starts a fresh room, bypassing a pinned development room.

^{Written for commit 6e9c9b4. Summary will update on new commits.}

[ColtenOuO]

Squash the commit and remove Co-authored-by: xxx in commit message.

Addressing review feedback doesn't require adding a new commit message.

[ZuoTso]

Squash the commit and remove Co-authored-by: xxx in commit message.

Addressing review feedback doesn't require adding a new commit message.

Squashed into "95a5fd3" and removed the "Co-authored-by" trailer.

[jserv]

Excluding only previous means the pool never shrinks: each reload is a fresh uniform draw from every other problem at those levels. A candidate who already knows some of them can keep reloading until one of those comes up, which is the "practice a familiar problem" case #185 is about. Each reload also sends out a new agent while the old room keeps its capacity slot. If that is acceptable for a self-assessment mode, say so in the limits paragraph of docs/interview-selection.md. Otherwise, a follow-up could store the draws in a bounded, signed set (or keep them on the server per account) so that reloading moves through the pool instead of re-rolling.

[ZuoTso]

This is an intentional limitation of the opt-in self-assessment mode. Excluding only the previous problem reduces immediate repetition but does not prevent deliberate re-rolling. Strict retains the existing agent cleanup policy, accepting temporary capacity overlap. Both trade-offs are now documented in docs/interview-selection.md.

[jserv]

Squashed into "95a5fd3" and removed the "Co-authored-by" trailer.

You don't have to make statements like this, as GitHub always tracks commits.