moonsize 0.4.0 — review evidence and verified public package

moonsize 0.4.0 closes the review-evidence and public-package gap. It includes the current parsing/graph diagnostics, independent gzip semantics, function/package diff, experiments, path explanations, multi-metric budgets and self-contained native HTML delivery. This pre-1.0 minor release includes breaking library signatures; see the migration guide.

• Exact source: 5f1452aad48012ed40f1b3673ee3e1ddf24197cd, main. Primary materials commit 63ce102, followed by a byte-preservation fix found in the clean export. No forced push/amend.
• Target Linux CI: all jobs successful, on this exact SHA: 336 native tests, 730 CLI scenarios, four source demos, real fresh-offline browser/A4 PDF regression and current-source budgets (pass 0, refusal 3).
• Public library: BigSaltyMan/moonsize@0.4.0. Install with moon add BigSaltyMan/moonsize@0.4.0. This does not install a standalone CLI; build the source with moon build --target native --release.
• Clean external source exported from the exact SHA passed check/test/build/fmt, identical API and pinned-resource generation checks. Audited package: 104 files, 1,030,262 bytes. Registry and downloaded public ZIP exactly match SHA-256 ec537f92634421fb60ff7375143a054f67f40bb40c8e2b62ac140607be13a60a. Caches, local plans, review attachments/proposal and unrelated images are excluded; implementation, generated ECharts and all associated licenses are included.
• A second new external consumer installed the exact public version without a local override and compiled/executed parse/stat/build, real return_call path evidence, diff, experiment, Options, and budget pass/exceeded 0/3. Reproduce with node scripts/verify-release-consumer.mjs 0.4.0 consumer-result.json. Raw result is attached.

Reviewer entry points

• One-page project description, response to the original concern, public evidence index, API migration.
• Human-reviewed example PDF, 5 pages, long-name/diagnostic PDF, 4 pages, all nine page reviews and original environment. These are byte-identical copies of the recorded historical Windows PDFs, not newly claimed clean-CI human review. CI generates its own automatically checked PDFs.
• Pinned resource/version/Apache+d3 licenses. ECharts 5.6.0 SHA-256 bf4a223524e40b77c304bec67e1222cf551f14880cf42c69dc046558e11c07b1; .gitattributes prevents Windows checkout/archive newline conversion. Native CLI is 2,893,824 B on Windows (prior inline-resource change +2,088,960 B); Wasm budget scope is unchanged.

Boundaries

Retained is dominated encoded function weight in a fixed graph/root model, not exact rebuilt savings or a general bound. Model-unreachable functions are investigation candidates. Paths are reference-model evidence, not execution traces. Independent gzip measures local compressibility, is non-additive and is not network attribution. Whole gzip is one representation under this compressor configuration, not HTTP/TLS or deployment cost. Source demos deliberately change optional output; base checks/rollback do not prove full behavior equivalence. Strict success does not prove exact savings or full Wasm validation.

Offline acceptance used copied HTML fulfilled only as an HTTP main document in a new context that was offline from creation, with zero external requests. A4 print, charts and state restoration passed. file:// OS double-click, non-Chromium and physical printers remain unverified. Library render_html's empty optional ECharts source retains CDN behavior; the native CLI embeds the pin. No new performance/scale promise or contest approval guarantee is made.

Attached JSON records supply source → exact CI → audited public ZIP → public consumer provenance; repository snapshots/PDFs remain directly readable even when CI artifacts expire. Toolchain: moon 0.1.20260915 / moonc v0.10.13+cbb11c36f; Node24.18.0 for local orchestration.