End-of-term attendance export, and stop question deletion from breaking reports

[percolator]

Two changes, both about answers surviving long enough to report on.

An end-of-term participation export

GET /api/reports/participation.csv?from=&to= — one CSV for a whole semester, reachable from the dashboard under End-of-term participation with an optional date range.

One row per student, one column per session, yes/no per cell. Correctness is deliberately absent: this is the participation-credit record, not a mark.

username,name,2026-01-14 Alignment (kq8fzm),2026-01-21 Assembly (p4wtrn),sessions_attended,sessions_total
adam,Adam A,yes,yes,2,2
bo,Bo B,yes,no (1/2),1,2

What "took part" means. The student answered both the pre and the post round of every question that was asked in both bouts. Two deliberate softenings:

• A question that never got its second bout counts against nobody, and a session where no question ran both bouts is blank rather than an absence — the teacher not running a post round is not the student's absence.
• A failing cell carries the fraction (no (1/2)), because a strict all-or-nothing verdict would otherwise make a student who answered three of four questions indistinguishable from one who never turned up.

If you'd rather have a threshold ("at least 75% of pairs") than all-or-nothing, the rule lives in one place — service.semester_participation — and the fraction is already in the JSON.

Personal data, so teacher-only, restricted to the requester's own sessions, Cache-Control: no-store, and labelled do-not-project like the Participants view.

Deleting a question silently broke reporting

Found while answering how long answers survive. Probing a real deletion against a session with a recorded answer:

delete question:              204
comparison after delete:      404
participation after delete:   AttributeError: 'NoneType' object has no attribute 'choices'
participation.csv:            AttributeError: 'NoneType' object has no attribute 'choices'
orphaned rounds: 1    surviving answers: 1

Nothing cascades from a question to the rounds that asked it. Deleting a used question did not remove its answers — it stranded them: the rows stayed in the database, Round.question became None, and participation_report dereferenced it, so /participation and /participation.csv raised for every session that had used that question. The data was intact; the only way to read it was gone. The trigger looks like tidying up a quiz.

Now refused with 409 pointing at the reset button, for when discarding answers really is the intent. The report also skips a round whose question is missing, so sessions already damaged by the old path become reportable again.

Tests

90 backend, 5 frontend unit, 9 browser specs.

• test_semester_report.py — both-bouts-not-correctness, every asked pair required, single-bout questions not held against anyone, date filtering, CSV shape, partial attendance visible, teacher-only, not public.
• test_answers_survive.py — deletion refused, unused question still deletable, report survives data damaged by the old build.
• semester-report.spec.mjs — drives a real lecture in a browser with two students (one answers both bouts, one only the first), downloads the CSV from the dashboard and checks what it says.

Not addressed

The database is still a single SQLite file with no backup. The app now protects answers from itself, not from the volume — exporting after a lecture remains the only real safeguard.

---

Generated by Claude Code