What Problem This Solves

Fixes #100944. Signal could silently drop an already-authorized inbound message when reply-session initialization lost its final optimistic concurrency race. The generic inbound debouncer logged the failed flush but did not retain the Signal payload needed for another delivery attempt.

Why This Change Was Made

The shared reply-session initializer already retries once under its writer lock, so another unbounded session-layer loop would hide sustained contention. Signal is the narrow owner that still has the original inbound payload, making a bounded channel retry the appropriate recovery boundary.

This revision:

• recognizes the exact conflict through nested cause and error wrappers;
• retries the original ordered batch after 1s, 2s, and 4s backoffs;
• keeps the debouncer's same-conversation execution slot reserved during backoff, preventing newer buffered text from overtaking the failed batch;
• binds retry delays to the Signal monitor abort signal and tracks retries in the monitor's drain set; and
• leaves non-conflict failures and exhausted retries terminal and logged.

User Impact

Signal follow-up messages that hit this transient session race are retried instead of being silently lost. No configuration or migration is required. Monitor shutdown cancels pending backoff and does not dispatch stale retries afterward.

Evidence

• corepack pnpm test extensions/signal/src/monitor/event-handler.reply-session-conflict.test.ts extensions/signal/src/monitor.tool-result.pairs-uuid-only-senders-uuid-allowlist-entry.test.ts on Blacksmith Testbox tbx_01kx591b6gg472mwcn1mhhhcqt: 11/11 passed across 2 files (run 29072172902).
• corepack pnpm check:changed on the same Testbox passed all guards, all four TypeScript lanes, and core lint. Extension lint found three local fixup issues; all were corrected and targeted extension oxlint is clean.
• Exact pushed-head GitHub CI run 29073336485 completed successfully with zero failed jobs.
• Focused regressions cover nested error graphs, bounded exhaustion, non-conflict failures, ordered batch retry, newer-message ordering during backoff, abort cancellation, and retry task draining.
• The production monitor boundary test emits an inbound Signal event through monitorSignalProvider, aborts the monitor during backoff, and proves no post-stop dispatch.
• Node and Bun both verified node:timers/promises abortable unref delay behavior (AbortError).
• oxfmt and git diff --check: clean.
• Fresh reviews accepted and fixed two lifecycle/ordering findings. Final full-branch Codex review on exact pushed head ef26709dcd74d632d07e6654b4f5870d7f555bc4 reported no actionable findings (confidence 0.82).

Live Signal account proof was not available in this checkout; the monitor-level test is the closest credential-free production-path proof.