beforeInsert overwrites a system-supplied last_update_at, so seeding a stalled task takes two passes instead of one #63

Found while establishing the seed-writes-history answer for #32. Filed unassigned for triage — the two-pass workaround is measured, documented in AGENTS.md and works today, so this is an ergonomics card, not a blocker for #7.

What happens

task.hook.ts's beforeInsert leg stamps unconditionally:

if (ctx.event === 'beforeInsert') {
  input.last_update_at = now;
  return;
}

That is exactly what #3 specified and it is right for every ordinary create. But lifecycle hooks do run on the seed path — the platform's SeedLoaderService.SEED_OPTIONS is { isSystem: true, skipTriggers: true, seedReplay: true }, and skipTriggers suppresses record-change automation, not hooks — so a seed's last_update_at is overwritten with the boot clock even under a system context.

Measured on @objectstack/runtime 17.2.0: a system-context insert carrying last_update_at: <45 days ago> stores now. completed_at is unaffected (the hook does not touch it on insert), so the done half of history seeds in one write; only the stalled half needs the extra pass.

Current workaround (shipped, documented)

A second seed dataset on the same object in mode: 'update', matched on externalId, carrying only last_update_at. The beforeUpdate leg deliberately does not stamp on an administrative write, so the value lands. See the "How to write history" section of AGENTS.md and test/seed-history.test.ts.

It works, and it costs every seed / import / fixture author an extra pass plus the knowledge of why.

What could change

The hook can see the context. Measured on the same boot, a beforeInsert context carries:

ctx.session      = { isSystem: true }
ctx.input.options = { context: { isSystem: true } }

So the insert leg could stamp only when the caller did not supply a value from a system context — roughly "a trusted writer replaying history may set it; everyone else gets now".

Not done here for two reasons, both worth weighing at triage:

• src/hooks/ was outside A task cannot be created directly in done — decide whether that is the intent #32's file surface, deliberately.
• It widens what a system caller may do to the stagnation signal, which is the one column the hook exists to protect. The current shape has the virtue that the ONLY way to move last_update_at backwards is an explicit, separate, visible administrative write. That may be the better design even though it costs a pass.

If the answer is "leave it", this issue should close as by-design with a line in the hook's header saying so, because the next seed author will ask the same question.

Acceptance

Either:

• the beforeInsert leg honours a system-context-supplied last_update_at, test/seed-history.test.ts's "an INSERT cannot carry last_update_at" assertion is inverted, and AGENTS.md drops the second pass; or
• the hook header records the deliberate refusal, and AGENTS.md keeps the two-pass shape as the permanent answer.