fix(task-board): a stale card repo must not veto binding the sole repo

[viktormarinho]

Regression from #7159, on the Super Agent dispatch path and not behind a flag — so it's live for every org.

pickSoleTaskRepo gained a preferredRepo parameter that filtered the choices, and an empty filter result meant "don't bind". But task.repo is free text (z.string()) that goes stale — a renamed repository, a value from the github-connection era, a typo. So an org with exactly one repo stopped binding it whenever a card named something else, falling through to the mid-run TASK_ADD_REPO pick and costing a turn. That org bound its sole repo before the parameter existed.

The preference is now a hint: it narrows when it matches, and is discarded when it doesn't.

Multi-repo behavior is unchanged — a preference that matches nothing, or matches ambiguously, still declines to bind and defers to the mid-run pick. The existing test asserting that (pickSoleTaskRepo(twoChoices, "other/repo") === null) still passes, and still for the right reason: the filter empties, the original two choices stand, and length !== 1 declines.

choices	task.repo	before #7159	after #7159	this PR
1	unset	bind	bind	bind
1	matches	bind	bind	bind
1	stale	bind	mid-run pick	bind
N	matches	mid-run pick	bind	bind
N	stale/ambiguous	mid-run pick	mid-run pick	mid-run pick

Deliberately unflagged, like the regression it repairs: flagging it off would leave the broken case live.

Validation: the new unit test fails on main (expect(pickSoleTaskRepo(sole, "acme/renamed-away")?.id).toBe("one") → received undefined) and passes here. Workspace type checks, lint (14 pre-existing warnings, 0 errors), knip, formatting, and 8,717 unit tests all pass.

---

Summary by cubic

Fixes a regression where a card's stale repo value prevented binding an org's sole repository, forcing a mid-run pick that costs a turn. The card's repo is now a hint: it narrows the choices when it matches and is discarded when it doesn't.

• Multi-repo behavior is unchanged — a preference that matches nothing or matches ambiguously still defers to the mid-run pick.
• The fix is intentionally unflagged because the regression is live on the Super Agent dispatch path for every org.
• Adds a unit test that fails on main and passes here.

^{Written for commit 1c6767e. Summary will update on new commits.}