Fix assistant-tailed requests and keep the tail cache breakpoint

[olosegres]

Summary

• End every Claude request on a user turn. The subscription endpoint rejects assistant-message prefill ("This model does not support assistant message prefill. The conversation must end with a user message."). Only opencode's exact Continue with your tasks. nudge was handled, so any other assistant-tailed request failed with a 400 — e.g. a resumed session whose last saved turn was the assistant's, or opencode's max-steps prefill. The nudge is still dropped; any other assistant tail now gets a minimal Continue. user turn appended, so the assistant content is kept.
• Keep the tail cache_control breakpoint. opencode puts its prompt-prefix breakpoint on the tail message's last part, and the bridge stripped cache_control from every message text block, so text-tailed requests (plain user turns, compaction requests) never read the cached prefix (cache_read 0) — only tool-tailed requests did. Earlier text breakpoints are still stripped and none is added, so the four-breakpoint cap still holds.

This also fixes the max-steps prefill failure #26 is after; regression tests cover both its string and text-block shapes.

Test plan

• npm test
• npm run typecheck
• With the plugin loaded in OpenCode, two plain text turns in one session: the first writes the prompt cache, the second reads it back (cache_read 0 before this change).