fix(task-board): don't cache "no preview yet" for a full window

[pedrofrxncx]

Problem

A PR card read while its deploy is still running has no preview URL yet — and both caches store that not-ready answer as a normal value, then serve it as a fresh hit: the reads cache for 55s, the assembled card for 30s, layered. So the preview link shows up minutes after GitHub actually has it.

Seen on ELEC-245: the VTEX FastStore preview (sfj-a14bf72--electroluxecfaststore.preview.vtex.app) was live on PR #305 while the task card still showed only "3/9 successful checks" and no preview button.

Change

A per-entry hit-window override on fetch and fetchOrPlaceholder, computed from the stored value, used in the two places that can be incomplete:

• the GET_PREVIEW_DEPLOYMENT read, when it carries no environmentUrl yet;
• the assembled card, while checksStatus === "pending" and no preview has been found.

Both go stale immediately, so the next poll revalidates. Revalidation is already detached from the request path, so this costs one background rebuild per poll on exactly the cards still missing something — and nothing on the rest. A card that has its preview, or whose CI settled without one, caches normally again (isAwaitingPreview).

No TTL constants changed; no behaviour change for complete cards.

Testing

pr-cache.test.ts — two new cases pinning that a not-ready value is never a hit inside the default window, and that the override stops applying once the value is complete. checks-status.test.ts — isAwaitingPreview truth table (preview found, CI passing, CI failing, no CI all cache normally).

bun test apps/api/src/tools/task-board/ → 465 pass. The 10 failures are ECONNREFUSED :5432 integration tests needing a local Postgres, unrelated. bun run fmt, bun run lint (0 errors), tsc --noEmit clean.

---

Summary by cubic

Fixes the task board caching "no preview yet" for the full window, so preview links now appear as soon as GitHub has them.

While a deploy is still running, both caches served the not-ready answer as a fresh hit — 55s for reads, 30s for the assembled card, layered — so the preview could show up minutes after GitHub had it (seen on ELEC-245).

What changed

• Adds a per-entry hit-window override on fetch and fetchOrPlaceholder, computed from the stored value.
• Applies it where a value can be incomplete: the GET_PREVIEW_DEPLOYMENT read with no environment URL yet, and the assembled card while checks are pending and no preview is found.
• Incomplete values go stale immediately, so the next poll revalidates; the detached rebuild costs one background pass per poll on those cards only.
• Cards that have a preview, or whose CI settled without one, cache normally again.
• Adds tests for the override and the isAwaitingPreview logic.

^{Written for commit 725152a. Summary will update on new commits.}