The S2.8 forward-reference in plan.md previously named the
out-of-process explicit-decompression work as 'PR 3 of 3 — NEXT'.
This PR closes that out: PR #24 IS that work. Update the entry to
reference PR #24 and capture the as-implemented scope (which paths,
which deferred to S3.1/S3.2 for R2.6.1 disk-RDB compressed encoding,
which already covered by PR #23's transient-view hook).

S2.8 (Read-path hook) is now fully landed across PRs #21, #22, #23,
and #24. Next on the @ikolomi track: S2.9, S2.10, S2.11. Next on the
@GilboaAWS track: S1.2 (training) — now safe to merge because S2.8
PR 3 means the forked-child paths (AOF rewrite, RDB save) no longer
panic when they encounter compressed values.

* [S2.8] Out-of-process explicit decompression (3/3)

PR 3 of 3 in the S2.8 split. Wires explicit objectGetUncompressedView
calls into the two paths that bypass lookupKey() and thus do not
benefit from PR #23's transient-view hook:

  - rioWriteBulkObject (aof.c) — AOF rewrite child iterates kvstore
    via kvstoreIteratorNext; AOF on-disk format is uncompressed RESP
    per R2.6.5.
  - rdbSaveStringObject (rdb.c) — same iteration pattern; per R2.6.5
    (AOF preamble within RDB) and R2.6.8 (full-sync replication RDB),
    uncompressed bytes are mandatory.

Both today panic on OBJ_ENCODING_COMPRESSED (rioWriteBulkObject:
serverPanic("Unknown string encoding"); rdbSaveStringObject:
serverAssertWithInfo on sdsEncodedObject). The crash is latent only
because no dictionary training has merged yet (S1.2 pending) — once
training lands, BGSAVE/BGREWRITEAOF on a compressed key crashes the
server. PR 3 is a prerequisite for S1.2 keeping unstable continuously
safe.

R2.6.1 (compressed disk RDB encoding via RDB_ENC_COMPRESSED + AUX
dict entries) is intentionally deferred to S3.1/S3.2 (@GilboaAWS).
Until that lands, the conservative choice for ALL RDB targets is
uncompressed, matching pre-feature behavior. Disk RDBs pay the
decompression cost once at save time and re-acquire compression at
load time via the post-S1.2 sweeper.

Dict lifetime safety in the forked child: registry is a fork-time
snapshot; DDicts are immutable; child uses its own copy of the
registry pointers. No interaction with the parent's QSBR generation
counters.

Replication feed (feedReplicationBufferWithObject) does NOT need
wiring — already verified to operate on argv arguments and synthetic
SELECT robjs only, never on kvstore values.

DUMP/RESTORE/MIGRATE (R2.6.7) go through lookupKey() which now
decompresses transparently per S2.8 PR 2; nothing to change.

Tests deferred to the post-S1.2 transparency Tcl harness — that's
where real compressed values exist and BGSAVE/BGREWRITEAOF
round-trips can be exercised end-to-end.

Verified locally: make -j2 -C src SERVER_CFLAGS=-Werror clean both
with BUILD_ZSTD=yes (default) and BUILD_ZSTD=no; compression Tcl
tests 10/10 pass.

* docs(planning): mark S2.8 fully complete via PR #24 (3/3)

The S2.8 forward-reference in plan.md previously named the
out-of-process explicit-decompression work as 'PR 3 of 3 — NEXT'.
This PR closes that out: PR #24 IS that work. Update the entry to
reference PR #24 and capture the as-implemented scope (which paths,
which deferred to S3.1/S3.2 for R2.6.1 disk-RDB compressed encoding,
which already covered by PR #23's transient-view hook).

S2.8 (Read-path hook) is now fully landed across PRs #21, #22, #23,
and #24. Next on the @ikolomi track: S2.9, S2.10, S2.11. Next on the
@GilboaAWS track: S1.2 (training) — now safe to merge because S2.8
PR 3 means the forked-child paths (AOF rewrite, RDB save) no longer
panic when they encounter compressed values.