research: stage frozen ARX6 codec, conformance tests, and production-baseline harness

[baanish]

Scope

Import the executable general ARX6 research prototype as a draft, with reproducible conformance checks and a benchmark entry point against the complete production auto-selector.

This is not viewer integration. No codec registry, default auto-selection, UI, historical decoder, shipped dictionary, dependency, or deployment configuration changes. The experimental #g1L... wire is not supported by the current viewer and must not be distributed as a working share link.

The second-pass numeric reconstruction-program / dependency-order experiments are intentionally excluded. Their specialist wins did not justify adding them to the general codec.

Included

• The frozen, browser-safe native-frame codec under experiments/arx6/src/: raw WTF-8 tuples; causal lexical/character-class contexts; residual syntax mixer; earlier match discovery; base64url framing; CRC32; immutable SHA-256-pinned copies of the existing three 16 KiB priors. The evaluated core files are copied unchanged.
• experiments/arx6/check.mjs: 16 dependency-free checks using the repository's existing assets. The same cases are registered in tests/arx6-core.test.ts for Vitest.
• experiments/arx6/bench.mts: a new corpus-driven comparison with the entire production encodeEnvelopeAsync auto pool, the real Markdown formatter, round-trip assertions, complete link lengths, and exact legacy fallback. It emits measurements/hashes rather than unsupported viewer links.
• Qualified historical results, source freeze record, reproduction instructions, and an explicit integration/deployment checklist in docs/arx6-research.md.

The Node-only runtime contains a laboratory copy of the ARX2 tuple mapping. Production integration should expose/reuse the canonical helpers rather than preserve this duplication.

What the earlier research actually established

Across 93 held-out software-text proxies from 47 disjoint source groups: 7.41% aggregate full-link savings with fallback; 7.36% standalone; 90/93 strict wins; 3 standalone regressions; worst regression 2.77%. Entire artifacts fitting 2,000 characters went from 75 to 76.

That historical baseline was the shorter of exact ARX5 entropy/base64url and ARX2 with Node Brotli q11, not the full production portfolio and not a byte-identical brotli-wasm validation. Do not describe those numbers as results from the newly added production-baseline runner.

The existing report fixture was a separate diagnostic, not held-out evidence: the whole link shrank 2,616 → 2,301 characters and still did not fit. Exhaustive complete-line prefixes increased fitting source text 5,867 → 7,158 characters at the same 1,996-character link length. The 22% capacity gain is fixture-specific.

Verification

Executed in the authoring environment:

• Original prototype suite: 20/20 groups passed, including legacy golden vectors and 80 seeded random round trips.
• New node experiments/arx6/check.mjs: 16/16 checks passed, including source hashes, every UTF-16 code unit, lone surrogates, exact whitespace/control preservation, prior identity/ownership, metadata, corruption/truncation, hostile lengths, and fallback.

Not executed there:

• Repository Vitest wrapper, lint, complete typecheck/build, and existing app suite.
• New full-production-baseline benchmark.
• Application/Worker integration and browser/live-chat transport checks.

The environment could run Node against the attached prototype but could not clone the repository or obtain its npm dependencies. The draft status is deliberate; local core checks are not being presented as full application CI.

Before production integration / enablement

• Reuse canonical tuple helpers; validate reconstructed envelopes through the normal schema/normalization path; reconcile fragment and decoded budgets. The frozen core's 8,192-character limit includes #, unlike the app's body-only budget.
• Implement explicitly versioned async encode/decode routing and opt-in emission; leave existing wire versions untouched.
• Add bounded Worker execution/cancellation and measure worst-case latency/peak memory. Async installation is not non-blocking compression.
• Preserve the complete old auto result byte-for-byte on ties/losses/unavailability and evaluate independent real-agent data.
• Complete Chromium/WebKit, asset loading/skew, UI, and actual Discord/WhatsApp paste/click tests; update protocol docs, examples, pickers, and agent skill together.

Reproduce

node experiments/arx6/check.mjs
npx vitest run tests/arx6-core.test.ts
node --import tsx experiments/arx6/bench.mts corpus.json > arx6-comparison.json

The corpus format is a JSON array of { "id": "...", "envelope": ... } samples satisfying the existing payload schema. Third-party corpus text and its large license bundle are not silently republished as MIT application fixtures.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}