Skip to content

REPORT.md

Latest commit

 

History

History
52 lines (36 loc) · 5.44 KB

REPORT.md

File metadata and controls

52 lines (36 loc) · 5.44 KB
Raw

Rust-mode multi-frame delta transport report

Mechanism confirmation

The adapter already built a tail-only body: ordinal resolution receives messages.slice(rawStart), encodeOpenCodeMessagesToCk encodes only that slice, and native_messages is also sliced at rawStart. The regression was after body construction: when buildPagedModuleTransformPayloads returned more than one frame, rust-mode-transform.ts discarded the delta, re-resolved and serialized the complete session, and paged the full body.

The module dispatches every request with transform-page fields to handle_transform_page_value. Non-final frames are only staged. The final frame runs assemble_transform_pages, and only the reassembled value reaches handle_transform_unpaged_value, where expand_transform_tail_delta applies the tail-delta fingerprint gate. The new module test sends a real two-frame delta body larger than 512 KiB and proves that frame 1 is staged and the reassembled body passes the gate.

Source inspection also exposed a second full-sync trigger. When a visible tail message was replaced, the adapter put the replacement-prefix fingerprint in tail_delta.after, but the module gate compares after with the previously acknowledged full-array fingerprint. The adapter now uses the prior full fingerprint for the acknowledgement gate while retaining the CK/native prefix fingerprints solely to compute the new full fingerprint incrementally.

The R11 corruption belt remains mandatory: every reused prefix message is still checked by prefixContentSnapshotsMatch before delta admission. A mismatch, missing cache, wire invalidation, forceFullWire, NEED_FULL_SYNC, or emergency/fold recovery sends a full wire. NEED_FULL_SYNC still retries once with a newly encoded full body. First process pass has no wire cache and is therefore full. HARD recovery is kept full by excluding emergency_recovery_armed passes from delta admission.

Measurements

Fixture: production-path Rust e2e harness, 1,000 persisted 1 KiB messages, then five small tail probes and one large tail whose delta requires multiple 512 KiB frames. Adapter time is pass elapsed minus the module-reported total. The before run used the original source; the after run used this change on the same machine.

Probe Revision Adapter p50 / elapsed Module Prefix guard State sync Wire build Transport Frames Wire messages / bytes
Small tail Before 256.7 ms 8.4 ms 0.0 ms 1.4 ms 26.6 ms 207.1 ms 7 full session; byte counter not present
Small tail After 178.5 ms 7.4 ms 0.0 ms 5.6 ms 0.2 ms 178.0 ms 1 3 / 5,227 B
Multi-frame tail Before 485.6 ms 29.8 ms 0.0 ms 6.6 ms 58.4 ms 447.6 ms 10 full session; byte counter not present
Multi-frame tail After 342.8 ms 24.7 ms 0.0 ms 4.9 ms 10.9 ms 348.7 ms 5 3 / 1,298,055 B
First-pass full control After 847.3 ms 304.6 ms 0.0 ms 15.8 ms 42.9 ms 1,058.5 ms 6 1,003 / 2,704,912 B

Structural result: steady wire construction is O(delta), falling from 1,000+ messages to 3-4 messages; a small tail is one 5-7 KiB frame; and the large tail remains a multi-frame delta instead of becoming the full session. The small-tail wire-build, state-sync, and prefix-guard budgets are met with substantial margin.

100 ms wall-clock status

The hermetic full-path fixture does not meet the revised 100 ms wall-clock target because its externally connected TCP provider route has a repeatable one-frame round-trip floor: the five post-fix 5-7 KiB samples spent 173.1-180.0 ms in transport while the module took 7.2-8.2 ms. The harness intentionally starts the daemon with modules: {} and connects ck-mc as an external provider, so this floor remains after serialization and frame count become constant-size. It is not session-size work: wire build is 0.2 ms, prefix guard rounds to 0.0 ms, and only one frame is sent.

The e2e test always enforces the structural properties and CPU-stage budgets. MC_RUST_E2E_STRICT_PERF=1 additionally enforces transport < 30 ms and adapter total < 100 ms on a transport substrate without the hermetic external-provider floor. The plugin-level 1,000-message acceptance test uses an immediate module client and enforces adapter elapsed < 100 ms directly.

Files

  • packages/plugin/src/hooks/magic-context/rust-mode-transform.ts
  • packages/plugin/src/hooks/magic-context/rust-mode-transform.test.ts
  • crates/mc-module/src/lib.rs
  • packages/e2e-tests/src/rust-harness.ts
  • packages/e2e-tests/tests/rust-multi-frame-delta-perf.test.ts
  • REPORT.md

Gates

  • cargo fmt --all -- --check — passed.
  • cargo test -p mc-module -p mc-store — passed (mc-module 685 passed / 3 ignored plus real-daemon 1 passed; mc-store 104 passed).
  • bun test in packages/plugin — passed (3,226 passed, 0 failed).
  • bunx tsc --noEmit in packages/plugin — passed.
  • bun run test:rust-e2e in packages/e2e-tests — passed (11 passed, 4 explicitly gated/skipped, 0 failed).
  • bunx tsc --noEmit in packages/e2e-tests — passed.
  • cargo test -p mc-module paged_authority_tail_delta_reassembles_before_full_sync_gate — passed.
  • bun test src/hooks/magic-context/rust-mode-transform.test.ts — passed (39 passed, 0 failed).
  • MC_RUST_E2E_STRICT_PERF=1 — not run as a passing gate on this hermetic external-provider transport; the measured 173-180 ms one-frame floor is documented above.