effects: the operation vocabulary is not node's - #1753

Groundwork for step 4 of emergent_testing/todo/share-browser-console-runner.md, found by looking at what step 4 actually requires: before a host-independent operation can move out of effects/node, the types it is declared in need a home that is not node's either.

What moved

OpResult, IoError, IoErrorInfo, IoChannel, IoResult, and the ioError / toIoError constructors, from effects/node to effects/ — beside NotImplemented, which OpResult is defined in terms of.

Nothing in any of them names a host. "The runner cannot dispatch this" and "the host tried and failed" are the two ways any operation goes wrong, on any host; normalizing a thrown value into serializable effect data is what any interpreter's catch does.

isNotFound stayed in effects/node, and the line between it and the two constructors is the useful part: it reads ENOENT, a POSIX filesystem code that a host without a filesystem never reports. Being about a host failure does not make a thing host-agnostic — being about no host in particular does.

It already had a victim

effects/memory/types.ts — typed key-value memory, no host at all — imported OpResult from ../node/types.ts. It now takes it from the layer it belongs to. That is what makes this separation of concerns with a consumer today (DESIGN.md §4), rather than an extraction on the promise of one.

The operations themselves (all, await, sandbox, and the unsettled fetch/import/now) deliberately did not move: until step 5's browser interpreter is a second implementer, moving them makes nothing shorter or clearer, and §4 says to extract at the second real consumer. The todo now records that split and why.

Reconciling the design records

Review found three documents that had already ruled on these questions, or that this change made stale. All three are settled in-place rather than left contradicting the code:

• effects/todo/node-module-layering.md on isNotFound — the issue's reasoning holds, and the change reverted to it.
• The same issue on IoResult — it said the move must not happen, calling the core "the wrong destination". That was reasoned without a second host: a browser interpreter cannot declare fetch or import without the alias existing somewhere shared, so "exactly a Node-layer contract" is what a second host falsifies. Its stated test — "not an effect constructor or combinator" — also does not describe the file it was applied to, where NotImplemented already lives and is neither. It never mentioned OpResult at all. The one-site fjs/media/type cleanup it proposed survives untouched as an open task: a pure consumer should not name an IO alias whichever module it lives in.
• Its "do not leave re-export shims behind" caveat now records why this move is the exception: a shim keeps a dead coupling alive, which is the goal for every move in its table (fjs/text/sgr should stop naming effects/node at all). Node's own operations are declared in these types, so re-exporting what it genuinely uses is not that.
• Now / Fetch / Import are unsettled, and both this issue and step 4 now say so instead of asserting opposite answers. Neither was written knowing what decides it — which operations the step-5 interpreter implements — so step 5 rules and updates both files in one change.
• effects/todo/io-effect-migration.md is done but deliberately kept and cited by two live documents, and said IoError lives in effects/node/types.ts "beside the operations it belongs to". It now records both what the migration did (true while node's were the only operations there were) and what has since overturned it.

Nothing else changes

effects/node re-exports every moved name, so the several dozen modules that reach for them through it are untouched, and an operation's declaration still reads as one vocabulary next to its failure constructor. Three doc corrections came with the move: IoChannel said an effect "fails the way node IO fails" (now host IO); Operation called OpResult/IoResult "node conveniences" (now: they read that way because of where they were declared); and Effect called IoChannel "the node standard of NotImplemented | IoError" (now the standard channel any host's IO answers in). toIoError's doc notes that a browser DOMException has no string code and so normalizes through the message branch — correctly, since there is no OS code to report.

The six proofs for toIoError moved with it into effects/proof.f.mjs; isNotFound's three stayed with it in effects/node/proof.f.mjs.

Verification

With origin/main merged in (which brought #1750): npx tsc clean, fjs t 3522/3522, npm run cov exit 0 (100% line/branch/function thresholds). The commits after that merge are documentation and changelog only, re-checked with tsc.

Changelog: the core module gains public exports even though no old import path changed, so this is an addition to record rather than a silent refactor.

Changelog:

• effects: OpResult, IoChannel, IoError, IoErrorInfo, IoResult and
  the ioError / toIoError constructors are now importable from the core
  module, not only through effects/node, which re-exports them unchanged

🤖 Generated with Claude Code

https://claude.ai/code/session_016PyLwDNkPQM1uD1Tg5ApBg