[zzylol]

Why

Review follow-ups to the lazily grown KLL item buffer that #242 brought to main (from #238), plus one doc fix on UnivMon::check_shape:

• Stream-path regression. A 50M-update stream ran about 10% slower on 0.4.0 than before. The lazy buffer moved the full-buffer path into a #[cold] make_room, which left compress_while_updating with a single call site in push_value. The compiler inlined it there. push_value grew to about 2.3 KB and stopped inlining into update's callers. nm on the two release builds shows push_value as its own 2,348-byte symbol on 0.4.0, and absent (inlined) before.
• A stale comment. The KLL ASAPv1 decoder says a decoded sketch's level array is identical to the source's. Since the buffer is sized to the retained items, only the boundaries relative to the first retained item match.
• UnivMon::check_shape's doc. More than MATRIX_MAX_ROWS rows pass the check and build in memory, but to_asapv1 refuses them. The doc did not say so.

What

• Kll::compress_while_updating is #[inline(never)]. It runs once per level-0 fill, so keeping it out of line costs nothing on the hot path, and push_value inlines again.
• src/sketches/kll/wire.rs: the decoder comment states what matches.
• UnivMon::check_shape doc: a pyramid with more than MATRIX_MAX_ROWS rows builds but cannot be encoded. Callers that need an encodable state must check the row count.
• CHANGELOG entry under Unreleased.

The other review items are already on main: the decoder's dead offset is gone (cursor starts at 0), and lazy_allocation_matches_eager_across_lifecycle compares clear(), decode-then-update and merge against the eager reference. The CHANGELOG blank-line and line-length nits were in #238's own text, which #242 did not carry over.

How

One attribute and doc text; no behavior change. Retained items, coin tosses and ASAPv1 bytes are unchanged; the existing eager-equality and golden tests cover them.

Before

nm --size-sort kll_bench (0.4.0):  0x92c  Kll::push_value   (out of line, compaction inlined)
stream, k = 469:                   34.7 ns/update

After

nm --size-sort kll_bench (this PR): Kll::push_value absent (inlined), Kll::compress_while_updating 0x84f
stream, k = 469:                    32.5 ns/update

Verification

Benchmark: a local example, not committed. It runs a 50M-update stream with xorshift input and with_seed(k, 8, 7), and creates 1,000,000 KLLs (k = 469) with 10 updates each, held 10,000 at a time. Release build, pinned to one core (taskset -c 50), the three builds alternated in each round, 7 rounds. The host was shared and its load swung from 22 to 70 during the session. Rounds 2–4 had the lowest load, and all three builds were stable within 2% in them; the numbers are their ranges:

	before #242 (60101c6)	main (eb04e7f)	this PR
stream, k = 469	31.1–31.5 ns/update	34.7	32.5–32.6
stream, k = 200	32.7–34.4	34.3–35.5	32.2–32.3
create 1M KLLs, k = 469, 10 updates each	13,890–14,546 ns/sketch	923–1,123	898–924

Remaining gap: within 4% of 60101c6 at k = 469 under that load (main: 10–11%), within noise at k = 200; a review run on a quieter host measured parity. The CHANGELOG uses the same figure. The lazy buffer's 15× creation speedup is kept.

Checks: cargo fmt --all -- --check; clippy -D warnings with --all-features, default features, octo-runtime and experimental; cargo test --all-features --locked (2,233 passed) and cargo test --locked (1,977 passed); RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --all-features --locked.