fix(sdk): make durable run recovery reject incomplete history

[Kaguya-19]

When durable event pages overlap or stop before the run's recorded last sequence, SDK recovery can duplicate output or report a partial result as completed. Recovery now advances only through contiguous positive event sequences, ignores replayed events, and returns the existing result_unknown state for gaps or incomplete history. It only queries the original run.

Extraction Inventory

Baseline: standalone_alpha1 at 70228cfed; selectively extracted from d56d5bcc6.

File/block	Change	Verification
packages/sdk/src/client.ts: readDurableRunEvents	Deduplicate replay; validate continuity; reject gap, invalid sequences and pages without progress	13 new regression scenarios
Same file: runs.result	Incomplete history returns result_unknown with original recovery reference	No submit_turn during recovery
packages/sdk/test/transport.test.ts	Page/cross-page duplicates, misleading nextSeq, gap, missing sequence, early EOF, no progress, invalid sequence	Existing paged output/usage/finishReason preserved

Validation

• Before implementation: 13 new cases failed; existing paged recovery passed.
• After implementation: all 14 focused runs.result cases passed.
• SDK typecheck and build passed.
• Full SDK transport file: 139 tests passed.
• Lockfile and SDK dependency manifests match the existing installation; tests used ignored dependency links in an isolated worktree.

Public SDK result types, result projection, AgentLoop and Gateway protocol are unchanged. Materials, skill selection, Lantay adapters/vendor and large acceptance data are excluded. This PR and the Docker extraction are independent, both targeting standalone_alpha1.