perf(sessions,think): memoise the auto-compaction estimate and drop per-turn config re-reads (#2226)

* perf(sessions): memoise the active path token total for auto-compaction

The auto-compaction threshold check ran a sized path walk after every
insert, so a turn's cost grew with the transcript. The core now keeps the
active path's total in memory: a tail append extends it, an update of a
counted row shifts it by the delta, and any write that changes the path or
its overlays (branch append, delete, clear, compaction, import) drops it so
the next read re-derives it from one walk.

Also: a fresh-id append probes existence by key instead of reading the
row's content, and Think skips the per-request think_config writes for
lastClientTools and lastBody when the serialized value is unchanged.

* perf(think): read the skills fingerprint from think_config once per object

* fix(sessions): bound the token memo at the path cap and drop it on a rolled-back cutover

The memoised path total kept every row's estimate past MAX_PATH_DEPTH,
while the uncached walk only sees the newest window; the memo now tracks
its depth and is dropped once a tail append would exceed the cap. The
sync write aperture gains abandon(), which forgets the tail and token
caches; Think and AIChat call it when the cutover transaction throws
after the message write, so a rolled-back row is not counted.