[LeadcodeDev]

WS4 of #521. Stacked on #529 (WS2) — review that one first; this PR's diff will shrink to its own commit once WS2 merges.

What changed

The canvas invented its trigger. buildInitialNodes created one node at a hard-coded __trigger__ id from two workflow-level props, and buildInitialEdges invented an edge to every root. Nothing about it came from the graph, so a workflow could hold exactly one trigger, and it could never be moved, deleted, or told apart from another.

Triggers are now drawn from graph.triggers — one node each, at its own graph-local id, positioned from the layout, draggable and deletable, with the same card, selection ring, error badge and + handle a connector node has. buildLayout records their position, which it used to discard. Nothing is invented any more, so TRIGGER_NODE_ID and the three helpers that built the fiction are gone.

Configuration moved from the workflow to the node. The panel edits one trigger and reports through onChange; it has no save button of its own, because a trigger is saved with the graph exactly like a connector's config. The context menu offers "Ajouter un déclencheur" in place of "Configurer le déclencheur" — the part of the original context-menu request (#520) that was never delivered.

A gap this closed on the way

projectGraphErrors routed errors by connector_id alone. The five GraphError variants that name a trigger — InertTrigger, EmptyEventTrigger, UnknownEventName, EdgeIntoTrigger, BranchedTriggerEdge — therefore landed in the graph-level banner instead of on the node at fault, which is precisely what #525 asks for. The backend has emitted trigger_id since WS1; the editor now reads it.

This was found because a test was written against the real pipeline and left failing rather than weakened to pass.

Two consequences worth naming

run-now-feature read the subscribed event names from the /trigger route that WS2 deletes. It now derives them from the graph's triggers.

The workflow list loses its "Déclenchement" column. It said "Manuel" or "Sur événement" — a label that cannot describe a workflow holding several triggers of different kinds — and the list response carries no graph to summarise. Restoring a summary needs a new API field, so it is a change of its own rather than a passenger on this one. The two tests asserting that column were deleted rather than contorted, since their intent is superseded.

Acceptance (#525)

• A second trigger can be added, configured and wired to its own connector.
• A trigger can be deleted, and its edges go with it.
• A trigger wired to nothing is refused on save, and the message lands on that trigger.
• A manual and an event trigger coexist and read differently on the canvas.
• The old workflow-level trigger panel is gone, with nothing left calling its hooks.

Verification

pnpm check	clean, 544 files
pnpm vitest run	1819 passed, 0 failed
pnpm exec tsc --noEmit	28 errors, identical set to main's baseline
pnpm build	clean

Closes #525

Follow-up from running it (second commit)

Two things made the node look real without behaving like one, both found by using the editor rather than by a test.

A trigger could not be wired by hand — the line vanished on release. isBranchDeclared refused every connection whose source was a trigger. That was correct while the canvas invented those edges and no user could draw them; it is wrong the moment a trigger is a node you add unwired. It now refuses only a branched edge leaving a trigger, which is the rule the domain actually states (BranchedTriggerEdge). Connector-to-connector was never affected — pinned by a new test alongside, so the distinction is recorded rather than assumed.

A trigger's kind is now chosen when adding it, not afterwards in its panel. Treating "manual or events" as a setting was the wrong model: it is what the node is. A connector picks its kind at creation and never changes it; a trigger is no different. The context menu offers "Manuel" and "Sur événement(s)" as two ways to add one, and the panel drops its mode tabs — it configures the events of an events trigger, and a manual trigger has nothing to configure, so adding one opens no panel at all.

The three panel tests covering mode switching were deleted rather than rewritten: their subject no longer exists.

Re-verified after both: pnpm check clean, 1819 frontend tests pass, tsc at main's 28-error baseline, build clean. Four unrelated failures in the full sweep (hr, planning, quotes) were load-induced timeouts on a loaded machine — all pass when re-run alone.

[codspeed]

Congrats! CodSpeed is installed 🎉

🆕 11 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• client_ip_x_forwarded_for: 4.6 µs
• client_ip_x_real_ip: 3.3 µs
• inject_headers_allowed: 4.5 µs
• inject_headers_blocked: 4.5 µs
• client_ip_fallback_only: 4.3 µs
• ip_v4_key: 4.2 µs
• decode_realistic_token: 37 µs
• extract_claims: 38.1 µs
• client_ip_forwarded_header: 4.2 µs
• ip_v6_key: 4.2 µs
• organization_key: 4.2 µs

---

[NathaelB]

🚀 Preview deployed: https://pr-530.mestier.fr

Synced revision <no value>.