fix(sessions): preempt paste-end Enter race with belt-and-suspenders double-Enter

[brianjones-v4n]

Summary

The Enter after bracketed-paste-end is occasionally eaten by a race with the paste-end (\e[201~) sequence on Claude Code v2.1.105+ TUIs. Today this is recovered asynchronously by verifyInjection, but in practice the recovery fires routinely (12 events in 48h on one production host) rather than rarely — meaning the "fallback" has become the de-facto primary path.

This adds a preemptive second Enter ~100ms after the primary Enter in the bracketed-paste branch of rawInject. The mechanism:

• If the first Enter was eaten by the race → the second submits the input.
• If the first Enter succeeded → the second arrives at an empty input prompt and is a no-op in the Claude Code TUI.

verifyInjection stays in place as the last-resort safety net for the rare case both Enters miss.

Test plan

• Added two unit tests in tests/unit/SessionManager-injection.test.ts:
  • asserts two send-keys Enter calls follow the paste-end sequence
  • asserts the inter-Enter sleep is present (~100ms)
  • asserts the verifyInjection doc acknowledges the new primary-path layering
• npx vitest run tests/unit/SessionManager-injection.test.ts — 8/8 passing
• Diff is contained to the multi-line bracketed-paste branch — single-line injection path untouched
• In-the-wild verification on the production host that produced the 12 events (will report back after merge)

Risk

Low. The second Enter is sent ~100ms after the first; if it ever fires when the input box is non-empty, that would be due to a separate ordering bug already present today — and verifyInjection would still cover the recovery.

🤖 Generated with Claude Code

[vercel]

@brianjones-v4n is attempting to deploy a commit to the sagemind Team on Vercel.

A member of the Team first needs to authorize it.