Let Almide call the store itself on the JS hosts through JSPI

[O6lvl4]

What changes

On the Wasm routes (Workers, the Node function adapters), Almide used to return step's reads and write to the JS host, and the host carried them out (adapters/step.mjs). Now Almide makes the storage calls itself, the same way the native host does:

• src/wasm.almd runs api.step to the end and calls two hooks, store_get / store_put. In Almide they are ordinary functions.
• The build passes --async-import store_get,store_put (--host js: async @extern via JSPI so Almide code can await JS I/O (prototype works on Node, workerd and the Cloudflare edge) almide#3353). The generated JS suspends the module through JSPI while each KV / Cloud Storage promise settles.
• serve returns a Promise. handle (no storage) stays synchronous.
• adapters/store.mjs (was step.mjs) only binds each hook to one store call. It also runs one serve at a time, so a suspended call cannot see another request's store.
• The Node requirement is now 24 or newer, which JSPI needs. All the function runtimes here are already Node 24.

Verification (2026-10-04)

Compiler: Almide fix-3353 at e9eb4bb90, built locally and passed as ALMIDE_BIN.

• npm test: 158/158 on Node 24.21.0
• Local workerd with a fresh KV state: 29/29. 20 concurrent POST /notes: all 20 stored.
• Deployed Worker with a freshly provisioned KV namespace: 29/29. wrangler tail showed 22 requests, all ok, and 0 exceptions.
  • 20 concurrent POST /notes across isolates kept only 5. This is the existing single-key read-modify-write limitation, now measured; this change does not affect it.
  • The Worker and the namespace were deleted afterwards.
• Cloud Run functions with GCS_BUCKET, via Terraform, in a disposable project (deleted afterwards), tested from a private in-VPC VM:
  • On the default nodejs24 runtime, every request answered 500. Google's image is Node 24.19.0, and JSPI is only on by default from 24.20.0; the generated init() refused with its JSPI message.
  • Added var.runtime / GCP_BASE_IMAGE (default nodejs24) and re-applied with nodejs26 (beta, 26.7.0). Results: 18/18 octet-stream, /notes 11/11, JSON showed the same 3 known framework rejections, and 403 without a token.
  • The bucket held the two notes, written by Almide through store_put.
  • 20 concurrent POSTs: 17 × 201 and 3 × 503. The 503s were Almide's storage_unavailable after GCS answered 429 to rapid writes of one object, and Almide logged that itself.
• CI moves to Node 24.21.0; engines is now >=24.20.0.

Draft until Almide ships --async-import

The pinned v0.66.0 release does not have the flag, so CI fails on this branch until a release includes almide/almide#3353. Then .almide-release and the checksums need bumping.

Found along the way:

• --host js marshal refusal hints "drop pub" for an unmarked fn; should suggest local fn almide#3360: the marshal refusal suggests "drop pub" for an unmarked fn; the actual fix is local fn.
• --async-import: JSPI needs Node >= 24.20, not >= 24 (refusal message and doc) almide#3362: the JSPI refusal and docs say Node ≥ 24, but JSPI needs ≥ 24.20.

🤖 Generated with Claude Code