fix(webhook): retry failed direct deliveries

[alexgunsberg]

Summary

• reserve deliver_only webhook IDs in a separate in-flight set so concurrent duplicates still send once even when a downstream send outlives the normal TTL
• commit successful IDs to the TTL cache only after downstream success
• release failed, raised, or cancelled reservations so the provider can retry the same ID

Regression coverage

• failed SendResult then same-ID retry succeeds
• downstream exception then same-ID retry succeeds
• cancellation releases the claim and re-raises, then same-ID retry succeeds
• two concurrent same-ID requests produce one downstream send, including after the normal TTL expires
• same-ID retry after a successful send remains suppressed

Verification

• pytest tests/gateway/test_webhook_deliver_only.py -o addopts= -q — 9 passed
• pytest tests/gateway/test_webhook_*.py -o addopts= -q — 54 passed
• focused Ruff, py_compile, and git diff --check — passed

Prior art and scope

Related to #47293 by @necoweb3. That older branch currently conflicts with main and also changes agent-mode completion semantics; this current-main patch preserves credit here while staying narrowly scoped to the deliver_only transaction required by the production watcher.

No live gateway, config, secret, webhook subscription, or deploy state was modified. Do not merge before the independent exact-head review of a0dc9d76bc6c4e3beb865e999ae82d109d06fa08 passes.

[alt-glitch]

This was generated by AI during triage.

Related to #47293: both allow a webhook delivery ID to retry after failure. This PR is narrowly limited to the deliver_only transaction, while #47293 also changes normal agent-mode completion semantics.

[alexgunsberg]

Independent exact-head review: PASS at a0dc9d76bc6c4e3beb865e999ae82d109d06fa08 against base 2446c8bb6755ff5e6feff4d26e425661edd4019b. Two independent read-only reviewers found no remaining P1/P2 after adversarial TTL/in-flight and stale-cleanup interleaving checks. Local evidence: 9 focused deliver-only tests passed; 54 webhook tests passed; a broader reviewer run reported 105 webhook-selected gateway tests passed with 7 skipped. Ruff, py_compile, and git diff --check passed. No merge performed.