Issue for this PR

N/A

Type of change

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

What does this PR do?

Adds experimental compaction.mode: "suffix" support to both session runtimes while keeping prepend as the default. Suffix compaction reuses the provider-bound request prefix, appends a bounded summary instruction, disables tool execution, and falls back once to prepend when the request is unsafe or the summary is invalid.

It also persists requested/used mode, fallback reason, duration, and token diagnostics, then shows them on the TUI compaction divider. Replacement plugin prompts retain their existing prepend semantics.

How did you verify your code works?

• bun run typecheck
• Core compaction tests: 14 passed
• OpenCode compaction tests: 62 passed, 1 skipped
• TUI tests: 196 passed, 1 skipped
• Schema tests: 17 passed
• Client generated check and typecheck
• JavaScript SDK regeneration and test
• Web build
• Linux x64 binary build and smoke test
• Human: Even though I used gpt5.6-sol-high to implement this, I have run it with a local Qwen3.8-27B-UD-Q3_K_XL.gguf via llama.cpp with different context sizes configured, and the suffix-based compaction works great :) I no longer have to wait for 90+ seconds on every compaction event.

Broader suites are not clean in this checkout: Core has two unrelated MoveSession git-apply failures; OpenCode has unrelated file-mode and response-timeout failures and exceeded the five-minute run limit.

Screenshots / recordings

Not included; the UI change is a text-only TUI diagnostic line.

Checklist

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