fix(session): use transcript position instead of lexical ID compare in prompt loop

[tiffanychum]

Issue for this PR

Closes #23490

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

SessionPrompt.run decided transcript ordering by comparing opaque message IDs:

lastUser.id < lastAssistant.id        // loop-exit check
m.info.id <= lastFinished.id          // post-finish reminder check

The HTTP API explicitly accepts caller-supplied messageID on prompt input, but those checks assume IDs always preserve OpenCode's internal monotonic ordering. When a caller supplies a messageID that lexically sorts greater than MessageID.ascending() output (e.g. msg_zzzzzzzz...), the comparison reverses and:

1. Loop-exit at finish="stop" no longer fires → loop runs an extra iteration after the assistant turn already completed.
2. The next request OpenCode sends ends with an assistant message → providers like Anthropic via OpenRouter reject it with "last message must be a user message".

The fix tracks lastUser/lastAssistant/lastFinished array indices during the same backward walk that already collects those messages, then uses array position (the canonical transcript order produced by MessageV2.filterCompactedEffect → page() → ORDER BY time_created, id) for both decisions. No new traversals; behavior is unchanged for OpenCode-generated monotonic IDs.

Two sites updated:

• Loop-exit check at the top of runLoop.
• Post-finish system-reminder loop after the user has sent additional messages.

How did you verify your code works?

Added a regression test in packages/opencode/test/session/prompt.test.ts:

it.live("loop exits when caller-supplied user messageID lex-sorts after assistant id", ...)

It seeds a session with:

• A user message whose ID is msg_zzzzzzzzzzzzzzzzzzzzzzzz (lex-greater than any hex-prefixed MessageID.ascending()),
• An auto-ID assistant message with finish: "stop" and time.created strictly greater so transcript ordering is unambiguous.

Then asserts prompt.loop returns the assistant message and the LLM mock receives 0 calls.

Verification:

• New test fails on dev (loop times out re-calling LLM with no replies queued — exactly the reporter's symptom).
• New test passes with the fix applied.
• Full session test suite green: bun test test/session/ → 315 pass / 0 fail.
• bun typecheck clean in packages/opencode.
• bun turbo typecheck clean repo-wide (13/13).

Screenshots / recordings

N/A — non-UI fix.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR