fix(data): recover queries after scheduler cancellation or rejection

[smiggleworth]

Clearing or rejecting a queued query previously left its refresh promise pending and prevented future refreshes from starting. Query starts now own a cancellable scheduling ticket that settles the current promise and releases queued ownership, including invalidation replacements and reconciliation retries. Existing generation tokens prevent obsolete queued work from settling a newer request.

The ticket's cancellation extension only resets internal bookkeeping and resolves a promise; scheduler clearing does not invoke subscribers. The redundant query startQueued flag is removed. Public signatures are preserved, including a Promise return when production admission silently rejects a promoted manual refresh.

Validation:

• Fourteen public query regressions pass: eight demonstrated red on the original implementation, and six review regressions demonstrated late-result publication after cancellation. Cancelling queued work now retires the aborted fetch generation, so late resolve/reject cannot replace the preserved snapshot.
• Formatting, lint/typecheck, build, unit (336), jsdom (1,619), Chromium (53), public types, and packed consumers (21) pass.
• Relevant tier2 query deduplication/abort benchmark passes (795 samples, 0.756 ms mean, 0.42% RME; no baseline speedup claim).
• All 60 repository checks, including architecture boundaries and the unchanged 256 KiB SSG hydration bundle limit, pass after rebasing onto root-lifecycle PR fix(boot): preserve reentrant root lifetimes during cleanup #398.

Cancellation limitation: clearing queued work preserves the last published query snapshot. If invalidation already aborted an in-flight request, that snapshot can still report refreshing until a later explicit refresh. This is documented and tested; cancellation does not publish a state change from inside scheduler queue clearing.