[kamescg]

Public identity packages and v1.1.0 SDK/CLI

Adds sixteen public ESM packages for canonical identifiers, classification/provider resolution, deterministic atom construction, protocol/periphery interactions, dynamic-fee math, deployment metadata, and SDK/CLI workflows. The release plan publishes fifteen changed packages to alpha; schema-org is already published and unchanged.

Core artifacts now match @0xintuition/contracts-v2@1.1.0-alpha.1, including the updated MigrationMode surface. Router artifacts match the separate @0xintuition/contracts-v2-periphery@1.1.0-alpha.0 package. Registry verification binds installed ABI/bytecode modules to exact lockfile integrities. release:check passes. Deployment compatibility and SDK mathematical parity require separate validation.

The public tree contains developer documentation, source, specifications, tests, and release tooling. Internal planning/session documents have been removed, public links updated, and npm publication checks reject assistant instructions and hidden working notes in tarballs. Existing public agent-facing SDK/CLI APIs remain available.

Validation

• Frozen install, contract registry verification, candidate artifact fingerprints, and release-version availability checks pass.
• Package/publication guards, provenance tampering tests, and supply-chain policy tests pass. Release-age exceptions bind only the two exact contract versions/integrities and expire November 2, 2026.
• Clean-room tarball build/install/import/type checks pass for all public packages, including the SDK and CLI.
• Conformance: 1,027 pass, 23 TODO, zero unexpected failures. The reference comparison still records 98 known differences; fixtures were not rewritten.
• Latest hosted CI status is shown in this PR's checks. No browser or live transaction validation is claimed.

Remaining before merging to main

• Integrate/review identity follow-ups feat(iid, iid-spec): floor weak account handles below anchor eligibility, with the spec amendment and conformance rule (P16g) #35, fix(iid-ladder): select identifiers like the reference on normalized evidence, with the alias-only and typed-wd filters pinned by a differential grid (P16h) #37, feat(classifications, iid): bring the catalogue, ladders and helpers to parity with the reference implementation and derive one podcast GUID per feed (P16d) #39, feat(iid-spec, iid, iid-ladder, iid-registry): register provider-local identifiers as class D with the reference's grammar, typing and P1 floor (P16f) #40, feat(primitives): bring the anchor builder to byte parity with the reference implementation, including the no-identifier fallback envelope (P05b) #41, and feat(primitives): accept a candidate IID on the anchor builder, export the reference's validators and the buildAtomAnchor name (P05c) #42; resolve or explicitly scope compatibility differences and missing capabilities.
• Establish SDK curve quote/revert parity against the published Solidity revision, including overflow boundaries.
• Validate packed-package acceptance and identity/schema migration behavior in Core v2 and the production application; confirm deployment compatibility and the URI-aware React scope.

Keep this umbrella PR in draft until those gates are satisfied. Passing artifact checks does not satisfy identity migration or consumer acceptance.

Getting started · Artifact provenance · Release guide · Conformance evidence

[jonathanprozzi]

Review at 5bb870e — not ready to publish or merge yet: 7 findings (6 Major, 1 Minor)

Scope: everything since my 19 Oct pass at 925a7b8, i.e. #45–#49 plus the atom-* alpha.1 reconcile (41 commits, 1,043 files), with the already-reviewed parity PRs #35–#44 treated as the base. Two independent passes: a conformance pass against the release plan, the package docs and the reference implementation's in-tree copies (both the commit the train README says it copied from and the reference's current tip), and an adversarial pass that executed reference-vs-train differentials and replayed the reference test suites against the built train packages. Everything below was reproduced by running the real built modules; the assembled gates were run natively afterwards (section at the end).

Findings

1. collectAtomIdentifiers drops provider-local identifiers, and the ported test now expects the loss.
packages/atom-classification/src/collect-atom-identifiers.ts:73 substitutes the normative ladder projector for the reference's application handle projector without binding providerLocalHandles: 'mint', so the projector rejects Spotify's unregistered scheme; :142 then overwrites the emitted identifierSet. Real built output for a MusicAlbum with identifier: int:spotify:album:4uLU… and hints [album, mbid, int:imdb:tt0133093]: reference [mbid, album, imdb], train [mbid, imdb]; with the album as the only identifier, []. The reference test at .port-source line 19 expects the album; the train's collect-atom-identifiers.test.ts:19 was rewritten to omit it. The enrichment port binds the policy correctly (atom-enrichment/src/identifier-rung-projection.ts:22), so this is a one-site omission. No README, CHANGELOG or conformance ruling declares it. Reference suite against the train: 1 pass, 1 fail.

2. Ambient credentials bypass the injected environment, and the fingerprint cannot tell.
createProcessingRuntime({ env: {}, cache: { provider: 'none' } }) with no ambient TMDB_API_KEY: the default TMDB plugin's supports() is false. Set only the process's ambient TMDB_API_KEY and it becomes true with the identical fingerprint processing-v1:895048…. The constructors fall back to process.env when the mapped key is undefined (atom-enrichment/src/plugins/providers/tmdb/index.ts:36, :162; same pattern in atom-classification/src/plugins/imdb/index.ts:167 and atom-enrichment/src/plugins/providers/youtube/index.ts:147), while atom-runtime/src/fingerprint.ts:46 hashes only config.env. The env-injection guards pass because they allow reads inside *FromEnv helpers; they do not check that ordinary constructors never call them. The pickProcessingEnv allow-list itself is complete (all 17 reference provider names forwarded; the ten filtered names are consumer controls consumed earlier).

3. Credentials survive in transport errors and nested causes.
Through the built fetchJson with six planted markers: HTTP 401, userinfo rejection and the engine's structured plugin-error projection are clean (47/47 redaction tests pass; engine.ts:591/:601). But a transport rejection carrying request diagnostics propagates unchanged from __shared__/http.ts:284 (markers 1–5 in message/stack, 1–6 on structured inspection), and a resolver failure is kept as cause (http.ts:455, constructor :126), where marker 6 survives. ProviderHttpError is on the public root surface (src/index.ts:146). Sanitise at the provider boundary including nested causes, with a regression through the resolver-failure path.

4. 279 dangling declaration maps ship in the four runtime packages, and the smoke gate exempts them.
declarationMap: true at tsconfig.build.json:9 in atom-parser (9 maps), atom-classification (99), atom-enrichment (164) and atom-runtime (7); npm-packlist selects them and each points at ../src/…, which is not published. scripts/smoke-tarballs.mjs:105 explicitly exempts atom-* from the map prohibition that 990a457 added for everyone else. Disable the maps and remove the exemption so a selected dangling map fails the gate.

5. Missing promtool now silently skips validation, including in CI.
The reference test required promtool when CI is set (.port-source pinned prometheus-exposition.test.ts:101); the train replaced that with a boolean-type assertion and skipIf(!promtoolAvailable()) on both substantive checks (pipeline-runtime/src/testing/prometheus-exposition.test.ts:101, :105, :116). The provisioning step from 8e614e1 is not in the current workflow (.github/workflows/ci.yml:32-41). Executed with CI=true and no binary: 10 pass, 0 fail, 2 skip. Provision it and fail closed in CI, or record the ruling.

6. Parser deviates from the reference pin without a recorded ruling.
Executed against the pinned reference modules in memory: parseAtom('did:example:123', {remoteFetch:false}) is url on the reference and plain_string on the train (atom-parser/src/detect.ts:215, HTTP(S)-only recognition); parseAtom('éé', {maxInputBytes:3}) is plain_string on the reference and throws INPUT_TOO_LARGE on the train (parse.ts:22, UTF-8 bytes vs UTF-16 length). Both are described in the README but not as parity rulings. Also stale: the README and migration runbook say DNS is not resolved or pinned (README.md:257, docs/atom-runtime-migration.md:94) while remote.ts:137/:258 now resolves, validates and pins an address. The bounded-read behaviour itself is sound: loopback, private, redirect-to-loopback, oversized, slow-body, data:/file:/ftp: all match the reference, and slow DNS is now bounded by the deadline where the reference was not.

7. Minor: the returned runtime exposes mutable preset factories after its fingerprint is fixed.
runtime.enrichment.presetFactories.default = () => [] takes the default inventory from 24 plugins to zero while runtime.fingerprint is unchanged (atom-runtime/src/runtime.ts:124, :133, :172). Freeze or snapshot the effective composition.

Also: three example workspace lock records disagree with their manifests (bun.lock:21, :32, :33 vs examples/atom-classification-plugin/package.json:24 and examples/hackathon-quickstart/package.json:16), which the registry validator does not cover because examples are not in the plan.

What checks out

All 20 release-plan entries match manifests and publishConfig.tag; internal ranges agree with bun.lock; publish order is valid; no workspace:/catalog:/file: in any selected manifest; React accepts protocol@4.0.0; CLI's 15 plan builders are function-identical aliases of the SDK's. The alpha.0→alpha.1 relabel is justified by the registry: curves, deployments, periphery already have 0.1.0-alpha.0 published; cli, iid, atom-parser are unpublished (404), as the release guide says. Focused conformance harness: 881 pass, 0 fail, 3 declared differences, 0 TODO (the PR body's "1,027 / 23 TODO / 98 differences" is historical; all 95 retired ids are still registered in the proof inventory and no conformance file changed since 925a7b8). The d8e665… default fingerprint reproduces; changing a TTL or preset changes it. Every one of 1,285 JavaScript export targets imports under Node 22 and Bun 1.3.11; no declaration import escapes its package. Replaying 102 reference suites against the built train: 1,196 pass, 49 fail, of which finding 1 is the only real divergence; the rest are reference-snapshot drift that the reference's own tip already matches (iTunes ISRC edge, MusicBrainz UA, TMDB query shape) or fixtures needing application files, servers or live transport. The earlier parity verdicts (#35, #37, #39, #40, #41, #43, #34) all still hold under fresh probes; #42's "75 agreements / 6 divergences" is superseded by 81 / 0 / 0. Supply-chain guard passes; the two release-age exceptions are bound to exact version, integrity and removal date 2026-11-02.

Consumer seams

The reference application's six SDK root imports and seven protocol root imports all exist on the built tree; primitives/anchor, identityLadderFor, getClassification(slug)?.atomCategory resolve. The Core service pins atom-parser/classification/enrichment at 0.1.0-alpha.0; this tree is alpha.1 after ef23b61 and the two are not interchangeable: parseAtom('int:isbn:…') is plain_string on alpha.0 and iid on alpha.1, NBSP handling flips, atomTypeSchema drops artist/music-album, rawInputKindSchema gains iid, the classification default preset grows from 16 to 27 plugins (adds int, Wikidata, Goodreads, Letterboxd, Places, OpenLibrary, Steam, CoinGecko, CoinMarketCap, Etsy, Shopify), and 31 + 47 root exports plus ~30 subpaths are added. MUSICBRAINZ_USER_AGENT is new in alpha.1. Nothing was removed. Core needs to adopt alpha.1 explicitly before "consumer acceptance" can be claimed.

Assembled gates run natively on this head

Run from a clean worktree at 5bb870e with a frozen install, Node 25 / Bun 1.3.11, after the sandboxed passes finished.

Gate	Result
release:check	pass
build	pass
typecheck	pass
check	pass
pack:dry-run	pass
smoke:tarballs	pass (clean-room install of every public alpha tarball)
guard:supply-chain	pass
test	pass with --continue --concurrency=4: 38/38 tasks, 5,634 tests, 0 failed, 2 skipped (the two promtool checks from finding 5). At full concurrency one run failed only pipeline-runtime provider-request-observation.provider-observer.test.ts ("measures the duration from dialed, not from begin", a 40 ms bound after a 5 ms sleep) and 33/33 pass in three isolated reruns, so that is a load-sensitive timing bound, not a product defect.
registry (npm view)	curves, deployments, periphery: 0.1.0-alpha.0 published; schema-org@0.1.0-alpha.0 published; cli, iid, atom-parser: not on the registry
publication guard probe	a planted dist/.codex/working-note.md fails the guard as intended; a planted dist/.env in the same tarball is accepted, because scripts/package-registry.mjs:71 matches agent directories and AGENTS/CLAUDE/GEMINI/SKILL.md only. Worth adding dotfile credentials (.env*, .npmrc) to that rule since files: ["dist"] selects anything under dist/.

Fix order: 1 (restore the reference assertion first, then bind the policy), 3, 2, 4, 5 as gate fixes, 6 as a recorded ruling. Happy to re-run the differentials and the reference-suite replay on the next head.