release: gate every extension on its own component round

[weilueluo]

Every one of this release's extension defects was found in a full tagged deployment plus an eight-case live canary, and every one was reachable far earlier: a Model export threw a plain Error and reached the kernel as a bare Wasm backtrace, temperature: null was forwarded to the provider, the OpenAI dialect never read generation.system_prompt, and the SSE decoder rejected every chunk because a host list<u8> does not arrive as a Uint8Array. All four are only reachable when the real componentized Wasm runs against the real host ABI, which nothing here did — the package tests exercise provider.mjs and index.mjs as plain modules.

Per-extension end-to-end smoke

tools/component-smoke.mjs drives one package's actual built components through a realistic round on Brain's own component-host worker, at the immutable Brain revision packages/env-aws-microvm/runtime/Cargo.toml already pins. That worker is a complete JSON-lines host over stdio with a capability seam, so the smoke needs no second host implementation and no Rust in this repository — .github/actions/component-host builds and caches the binary by that revision.

kind	round
Model	start → observe with a 2xx first chunk, a second chunk whose status is absent, then done → acknowledge; then a 500 first chunk
Agentloop	session_start with a hydration tail, then a message activation driving model_stream/journal_append/turn_finish; then a refused round
Environment	resolve → submit with a bundle → observe running → observe completed → acknowledge → release; then a refused dispatch
Tool	invoke under its declared grant against the package's own descriptor and bundle; then a refused grant

Each kind asserts that a failing export surfaces a typed extension-error carrying its reason, never a trap. @aexhq/model and @aexhq/agentloop publish no component, so their smoke compiles a fixture through the package's own public authoring path and drives that.

Reverting each of the four fixes and rebuilding shows the smoke catches all four, including the exact production symptom for the first:

temperature reached the provider
the sealed instructions never reached the provider
The input did not match /status 500/. Input: 'error while executing at wasm backtrace: ...'
Error: SSE chunks must be a byte sequence

Per-package staging and promotion

npm-publish.yml packed and published all nine workspaces as one lot. Now tools/npm-release.mjs records each package's intra-release edges in the manifest and derives the staging waves from them — no list to maintain, and a chain deeper than two waves fails closed rather than publishing a dependent before its dependency is visible. Staging fans out per package (@aexhq/model before @aexhq/model-openai, @aexhq/agentloop before the loops), each staged version is installed from the registry and smoked against its exact dist.integrity, and promotion moves each package on its own receipt for that exact archive, holding an unproven one on next instead of blocking the rest.

Every existing gate is unchanged: the release/sha-<40hex> tag and expected_commit check, the released-version integrity rule, the version-ownership source rule, the stage_run_id promotion consumes, and exact cross-release dependency verification. No package version changes and no file under packages/ is touched.

[weilueluo]

Run unchanged against the builds actually published to npm, the smoke rejects every one that shipped a defect and accepts only the current one:

published build	smoke result
@aexhq/model-openai@0.1.0	temperature reached the provider
@aexhq/model-openai@0.1.1	temperature reached the provider
@aexhq/model-openai@0.1.2	SSE chunks must be Uint8Array values
@aexhq/model-openai@0.1.3	passes

Reverting the remaining two fixes in the working tree and rebuilding reproduces the other two, including the exact production symptom:

the sealed instructions never reached the provider
The input did not match /status 500/. Input: 'error while executing at wasm backtrace: ...'