fix(ux): close 3 parcours bugs (Step 7 art. 25 flip, prohibition short-circuit, copy)

[abk1969]

Summary

Closes 3 bugs identified by the UX/parcours audit on the 7-step flow. Stacks on PR #7.

Bug #1 (CRITICAL) — Step 7 unreachable for art. 25 flip path

Symptom: An integrator who declares `nature='systeme_sur_gpai'` + `substantialModification='oui'` is reclassified as a GPAI provider by the classifier (PR #3, Item D), but Step 7 only showed the systemic-risk question when `nature==='gpai'`. Their `gpaiSystemic` stayed `null` → `isGPAI_RS` evaluated to `false` → they were classified as plain GPAI even at ≥ 10²⁵ FLOPs.

Fix: Step 7's visibility predicate now matches both paths:

```js
const gpaiQuestionApplicable =
answers.nature === 'gpai' ||
(answers.nature === 'systeme_sur_gpai' && answers.substantialModification === 'oui');
```

`canNext` updated symmetrically. The classifier already supported `GPAI_RS` for the flip path (PR #3 D1 tests pass) — this UI fix unblocks that codepath.

Bug #2 (IMPORTANT) — Prohibition short-circuit ignored carve-out claims

Symptom: `goToResultIfProhibited` jumped straight to verdict on any prohibition selection. A user who picks prohibition (h) AND claims the art. 5(2)–(3) carve-out (PR #3, Item A) would skip Annex I/III/50/GPAI entirely and land on RISQUE_MINIMAL with just the carve-out trace — even if they should have ticked Annex III §1 biometrics.

Fix: short-circuit only fires if at least one selected prohibition has no claimed-and-matched carve-out:

```js
const hasUnCarvedOutProhibition = prohibitions.some(id => {
const carveOutClaimed = !!carveOuts[id];
const carveOutExists = ART5_CARVEOUTS.some(c => c.appliesTo === id);
return !(carveOutClaimed && carveOutExists);
});
```

Bug #3 (MINOR — copy) — q7Sub / q7NotApplicable contradicted flipped integrators

Symptom: The "not applicable" message read "you are not a GPAI" — false for someone who declared substantial modification (PR #3 Item D made them a GPAI provider for the modified model).

Fix: Updated EN+FR copy on `UI.q7Sub` and `UI.q7NotApplicable` to acknowledge the art. 25 flip path. With Bug #1 fixed, this message is now reserved for users who are neither native GPAI providers nor substantially-modifying integrators.

UX note — Cumulativity rappel on Step 5

Added a one-line reminder to `q5Sub` (EN+FR) that art. 5 prohibitions, Annex III, and art. 50 transparency are cumulative — same underlying tech (biometric ID, emotion recognition) may need ticking in multiple steps because each tests a distinct legal regime. Reduces user confusion / risk of underdeclaring.

Tests + build

• 113 tests pass (unchanged count — no logic changes in classify.js; UI-only fixes)
• Build green: 279.21 kB JS / 84.52 kB gzip (+1 kB from new copy)
• Component smoke tests still pass (App still renders, Step 1 still navigates)

Test plan

• Integrator path: nature='systeme_sur_gpai' + substantialModification='oui' + Step 7 'oui' → verdict primary === 'GPAI_RS' (was 'GPAI')
• Carve-out path: prohibition (h) ticked + art. 5(2)-(3) carve-out claimed → user continues to Step 4, not jumped to verdict
• Non-flipped integrator (substantialModification='non') reaches Step 7 and sees the "not applicable" message with the new wording
• French UI: q7Sub mentions "art. 25" and "modèle GPAI tiers"
• Step 5 sub-question copy reminds user that regimes are cumulative

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai_act_compass	Ready	Preview, Comment	May 15, 2026 3:30pm