Read cached chats while other agents stream (#7074)

* perf(app): overlay pending replica rows on read instead of forcing persists

readRows flushed every pending change to storage, read, and retried
whenever the host revision moved. Every stream reducer flush (48 ms) and
every agent delta advanced that revision, so opening another agent or
workspace on a host with a streaming agent re-serialized the whole
pending batch per iteration and could spin until the stream paused.

Reads now run on the write queue, so no drained batch is mid-apply while
a read runs, and rows still pending in memory are overlaid for the
requested keys. A read never forces a persist and never waits for the
host to go quiet.

Persist retries after a storage failure back off exponentially (1 s to
60 s) and log the failure instead of re-serializing every second.

Before: one readAgent during a 200-commit stream cost 201 store reads and
200 forced persists; a permanently failing store was written 30 times in
30 s. After: 1 read, 0 persists; 5 writes in 30 s.

Fixes #4813
Refs #4814

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(app): log replica persist retry failures

---------

Co-authored-by: Ivan Malison <IvanMalison@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <paseo-bot[bot]@users.noreply.github.com>

docs/data-model.md

@@ -589,7 +589,10 @@ host, kinds, and ids. Opening a cached workspace uses exact workspace and projec
 directory scan. One invalid row is deleted and returned as a miss without affecting other rows.
 An invalid directory row and its affected checkpoint cursor are repaired in one transaction, so a
 later launch cannot accept a checkpoint for a partial baseline. Directory changes and their
-checkpoint are also applied in one transaction.
+checkpoint are also applied in one transaction. Reads wait for accepted changes to reach storage,
+then retry only when a requested row changes during the read. Unrelated agent streams do not
+invalidate the read. Directory baseline replacements and host removal or renaming still invalidate
+the affected reads.
 
 The cache is capped at 32 MiB and evicts whole hosts in least-recently-written order. Budget
 bookkeeping may scan opaque row sizes during a deferred write, never during host registry startup or