perf: reduce nested worker message serialization overhead - #200

Nested session metadata makes Worker.postMessage() miss the flat-object fast path. The general serializer then atomizes value strings, after already paying for cross-thread copies of the rejected primitive prefix. This change uses content-keyed string pooling for in-process messages and checks flat-object eligibility before allocating those copies. Successful flat objects keep the existing isolation path and reserve their exact property count.

Storage and cross-process callers retain canonical Identifier encoding. The pool remains per serialization, with no retained cache, wire-version change, or change to message scheduling.

Paired benchmark

Three separate quiet, on-demand c7i.4xlarge instances (16 vCPUs), Node 26.12.1, with the same verified release-profile candidate executable. Ten interleaved triples per instance; 25,000 warmup and 250,000 measured operations in 1,000-message batches. Only the synchronous send loop is timed; acknowledgments and checksum validation are outside it. Baseline: openclaw-v1.4.4-20261019-ef37c31de3-webkit-a2976f02e6.

The candidate reduces median paired sender cost by 30.6–34.3% and is near Node's median on each instance. This is a sender-latency result, not an equivalent end-to-end throughput claim: first-instance batch elapsed medians were baseline 1,069.93 ms, candidate 1,040.65 ms, Node 889.06 ms.

All thirty interleaved triples (ns/op):

bench/postMessage/postMessage-session-entry.mjs preserves the payload, counts, and sender-timer boundary for reproduction:

node bench/postMessage/postMessage-session-entry.mjs
./build/release/bun bench/postMessage/postMessage-session-entry.mjs

Regression guard

Thirty paired samples per cell, ten on each of the three instances. The decision statistic is the median of all thirty paired ratios, with a fixed ceiling of 1.03. Timed JavaScript, payloads and payload-specific operation counts are unchanged.

Supplemental flat-object pooled ratios: empty 0.989340; 16-key flat object 1.018603; 24 flattened leaves 0.991893 (30 pairs each).

Validation

• Native release: 446 passing tests across structured-clone-fastpath, workers/structured-clone, and workers/message-port-pipe; zero failures. Three existing tests are debug/ASAN-gated and do not run in release builds; no skips were added or changed.
• Separate Node/Bun oracle: 14 cases passed, covering cycles, shared references, indexed keys, null prototypes, nested getter mutation, and clone errors.
• Wire corpus: 288/288 byte-identical outputs to baseline, with Worker and MessagePort round trips and cross-deserialization in both directions.
• All 90 session samples and 720 microbenchmark samples completed with correct counts/checksums and natural cleanup. No timing samples were excluded or rerun to obtain a pass.
• Native CI: Linux x64 and Darwin arm64 passed for head 264b8a6d24e5007b98d1a3c16e1174eabd24bbd5. The CI merge commit has the identical Git tree to that head. Formatting, source lints and JavaScript lint also passed.
• Codex P2 autoreview: scoped-clean for the complete revised candidate.
• Upstream search found no existing fix for this nested-object serialization cost.