compaction: hold the next pass when the last one bought no room

[justrach]

Fixes #1686.

Root cause

An automatic compaction that left the context over the compaction point fired again on the next step. That happens when the provider's compacted state, or a summary that must keep a long live turn verbatim, can't get under the line. Every step then paid a compaction, and the model reread what it had just lost. The existing stall guard (#706) only catches repeated cuts at the same history boundary. A compaction that "succeeds" but leaves the context over the line rewrites history, so the guard reset on every pass.

Fix

• New compact_backoff.zig. If the pre-send gate is still over the line within one step of a history rewrite, the next compaction is held until the context grows by a twentieth of the window.
• The 95% wall always compacts, so recovery may trim as before. Real relief (the context drops under the line) ends the hold.
• The in-stream server compaction directive uses the same raised threshold while holding, so the server doesn't compact on every request either.
• A held pass is noted in the trace, and outside JSON mode it prints one line.

Tests

• compact_backoff: hold after a no-room compaction, release on growth, the 95% wall and real relief end a hold, and a compaction that bought several steps is not treated as a stall.
• Compaction, directive and turn unit tests: 250 pass.
• Tier-1 pre-push gate is green.