ducklink: update to v4.2.0

[zacharywhitley]

Six perf batches + a cold-start UX fix + an optional binary-size feature since v4.1.1. No API breaks, no WIT contract changes — additive minor bump.

User-visible wins

Measured on darwin/aarch64 with lto=true, strip=true:

Metric	Before	After	Δ
Aggregate 1M rows (sample_sum(v) FROM range(1M))	82 ms	38 ms	-54%
Parallel scalar (2-thread cross-extension)	12.9 ms	5.75 ms	-55%
Serial scalar dispatch	42 ms	11.2 ms	-73%
Cold-start on offline / bad-DNS	15 s	3 s	-80%
Binary size (offline build variant)	25.9 MB	21.7 MB	-4 MB

What's new architecturally

Dispatch (F/G/H/I/L batches)

• Column-native aggregate UPDATE: AggState is per-column typed accumulators picked at agg_init. Ungrouped chunks bulk-copy the whole input via extend_from_slice. Eliminates ~10M outer Vec allocs + boxed enum variants per 10M-row aggregate.
• CallbackRegistry Mutex → RwLock; Weak<Instance> on CallbackEntry so dispatch prologue is one atomic upgrade instead of HashMap<name, _>::get + Arc clone.
• Per-thread SCALAR_ARGS_SCRATCH: refill_colvec reuses Colvec buffers across chunks; TEXT/BLOB reuse via String::clear + push_str.
• write_colvec skips a redundant 2048-byte null-mask scan.

Discovery / docs (F1, F2, G4)

• Arc<CachedDocs> with generation-counter invalidation for the docs pipeline. render_help, ducklink.docs, and ducklink_search no longer rebuild ~639 DocRows per query.
• flat_vector handles hoisted out of every discovery TF's row loop.

Load-time (H4, J3)

• Base component-linker template cached per Engine: ~25 add_to_linker calls run ONCE per process instead of once per component load.
• Wasm bytes read ONCE per load and shared between Component::from_binary and parse_docs_from_bytes.
• wasmtime::Config::memory_may_move(false) (K1): 1-2% real per-row gain via base-pointer LICM.

Cold-start UX (J1)

The single most user-visible slow path in the codebase: opening DuckDB on hotel-WiFi / offline, running SELECT * FROM ducklink.modules, and staring at a black cursor for 15 seconds while a network fetch times out before the bundled snapshot loads.

• Catalog HTTP fetch timeout: 15 s → 3 s. The graceful fallback (bundled snapshot) is already in the binary.
• Async prewarm: Engine2::new spawns a background thread that populates the catalog OnceLock while DuckDB is still finishing extension setup. By the time the first user query arrives the catalog is usually already resolved. Best-effort; races are resolved via OnceLock::get_or_init's barrier.

Optional binary-size feature (J2)

New default-enabled network feature gates reqwest + rustls + webpki-roots + sha2:

cargo build --release                                     # 25.9 MB (with network)
cargo build --release --no-default-features \
            --features loadable,advanced                  # 21.7 MB (offline)

The offline variant is for air-gapped / minimum-size deployments; it serves only names in the bundled snapshot and refuses to download uncached blobs.

Community-extensions CI runs cargo build --release (default features) so this PR's distributed artifact keeps full network support — no change for end users of the CI-built extension.

Retracted design docs

docs/perf-ceiling-measurement.md, docs/wit-streaming-scalar-dispatch.md (RETRACTED), docs/wit-shared-memory-result.md (CANCELLED), and docs/guest-simd-investigation.md capture the measurement work behind this release and the reasoning behind rejecting several proposed architectural changes (~30% streaming projection turned out to be ~1.3%; ~27% zero-copy projection turned out to be <1%). New benches (streaming_validation, memcpy_floor) let anyone reproduce the findings.

PR history

Supersedes v4.1.1 (merged as #2201). All changes are on top of v4.1.1.

[zacharywhitley]

Superseded by #2214 (v4.3.0).