Deduplicate durable webhook entries before dispatch

[trakhimenok]

What changed

• Add an explicit DurableWebhookEntry capability for stable provider delivery IDs.
• Claim one leased inbox record per provider entry, process all sibling inputs under that lease, then complete once.
• Fail the lease when input processing fails or panics; live lease contention returns a generic 503 so the provider retries.
• Respect the HTTP envelope commit boundary: a 503 is written only before a platform response commits; committed responses are never appended or double-written.
• Treat nil, empty, and conventional zero IDs as “no durable ID” for legacy adapters.
• Persist only normalized safe failure classes through the store contract.

Why

Webhook delivery is at least once. Deduplicating per input would incorrectly skip sibling inputs in a provider batch; guessing durable IDs from any can collapse unrelated legacy updates.

Validation

• go test ./...
• go test -race ./...
• go vet ./...
• Router-backed multi-input claim/complete/fail lifecycle regressions
• Pre-commit 503 and post-commit no-double-write retry regressions

Dependencies

Stacked directly on #88 (agent/sanitize-webhook-panic-response). Also depends on bots-go-framework/bots-fw-store#9; the DALgo adapter is bots-go-framework/bots-fw-store-dalgo#4. Telegram #152 and the Sneat-Go adopter follow. Draft only; do not merge out of order.