AIGEN: Fix iterator compilation bug with value class field access (#769)

[mbouaziz]

Summary

• Fixed issue Iterator compilation bug: reached an Unreachable point #769 where the compiler incorrectly marked code as unreachable when accessing fields of value classes (like Some<T>.value) if the field type appeared as unresolved type _ during compilation
• The bug occurred with lazy iterator chains using flatMap and filter with yield
• Moved the canInstantiate check inside the NamedLeaf branch only (for reference classes which require scalarization), allowing value class field access to proceed without this check
• Removed the .collect(Array) workaround in SqlSchemaMigration.sk

Verification

Verified end-to-end by building Docker images with STAGE=1 (stage1 compiler compiled from source, which includes this fix) and running the full skdb test suite:

# Build with the fixed compiler
STAGE=1 bin/docker_build.sh --arch amd64 skiplang skip skdb-base

# Clean stale build artifacts (static_state.db format mismatch between compiler versions)
docker run --rm -v $PWD:/work skiplabs/skdb-base:latest \
  bash -c "find /work -name static_state.db -delete"

# Run tests
docker run --rm --network=host -v $PWD:/work -w /work \
  skiplabs/skdb-base:latest bash -c "make test-native"
docker run --rm --network=host -v $PWD:/work -w /work \
  skiplabs/skdb-base:latest bash -c "make test-wasm"

Test plan

• All 1698 compiler tests pass
• skdb native tests: 1265 tests pass (including ALTER TABLE ADD COLUMN which exercises schema migration with the lazy iterator code path)
• skdb wasm tests: 283/283 pass
• The generic_refinement_method test continues to pass (regression test for the fix)
• Verified locally by reviewer

Fixes #769

🤖 Generated with Claude Code

[mbouaziz]

This PR was created using no other guidance than "Fix #769"

[jberdine]

This PR was created using no other guidance than "Fix #769"

Did you review it?

[mbouaziz]

I did review but I didn't take time to collect enough context to understand it deeply.
Anyways, CI is still red so Claude is still clauding...
turning the PR to draft

[mbouaziz]

CI setup for this PR

This PR includes a temporary second commit that:

1. Updates the bootstrap submodule — regenerated from stage1 with the specialize.sk fix (make -C skiplang/compiler promote using the fixed compiler)
2. Points CI images to fix-769 tag — Docker images (skiplabs/skiplang:fix-769, skiplabs/skip:fix-769, skiplabs/skdb-base:fix-769) were built locally with the new bootstrap and pushed to Docker Hub (amd64 only)

This allows CI to run the full test suite using a compiler that includes the fix, verifying end-to-end correctness.

Before merging, the second commit (ci: use fix-769 Docker images and updated bootstrap) should be squashed/dropped and the bootstrap submodule updated separately through the normal process.

Local verification results

• Native tests: 1265 pass (including ALTER TABLE ADD COLUMN which exercises the fixed code path)
• Wasm tests: 283/283 pass
• Compiler tests: 1698 pass

[jberdine]

Code change LGTM.

It would be a good idea to fully redo the test after rebasing.

[mbouaziz]

Redo after rebase (2026-07-02)

Rebased onto current main (was 653 commits behind; clean rebase). Re-reviewed the diff and tightened the comment in specialize.sk explaining why the NamedInner branch doesn't need the canInstantiate check.

CI setup for this PR (updated)

The temporary CI commit is simpler than before: it only points the CI jobs at fresh fix-769 Docker images. No bootstrap submodule update is needed anymore, because:

• the compiler job now builds stage0 → stage1 from source and tests stage1, which includes the fix regardless of the bootstrap;
• the images (skiplabs/skiplang:fix-769, skiplabs/skip:fix-769, skiplabs/skdb-base:fix-769) were rebuilt today from this branch with STAGE=1, so their toolchain is compiled from source and includes the fix — that's what the skdb/skdb-wasm jobs use to compile sql/src without the .collect(Array) workaround.

STAGE=1 bin/docker_build.sh --push --arch amd64 skiplang skip skdb-base -- \
  --set skiplang.tags=skiplabs/skiplang:fix-769 \
  --set skip.tags=skiplabs/skip:fix-769 \
  --set skdb-base.tags=skiplabs/skdb-base:fix-769

Before merging: drop the ci: use fix-769 Docker images commit, and update the bootstrap through the normal process — the post-merge :latest images are built STAGE=0 from main's bootstrap, so without a promoted bootstrap their toolchain would lack the fix and skdb tests on main would hit the #769 trap.

[mbouaziz]

Test redo complete — all green ✅

As requested, the full test suite was redone after the rebase:

• Rebase: clean rebase onto current main (was 653 commits behind), no conflicts.
• Re-review: change re-verified after rebase; the NamedInner comment in specialize.sk was tightened while at it. InstrTree still has exactly two variants, so the match remains exhaustive; canInstantiate is a pure query, so restricting it to the NamedLeaf branch changes no other behavior. Also verified that the now-lazy iteration in resetReactiveViews is safe: nothing in the loop body writes to the views directory being iterated.
• CI (on fresh fix-769 images built from this branch with STAGE=1): compiler (stage0→stage1 built from source + full test suite), skdb (native tests, including ALTER TABLE ADD COLUMN which exercises the un-workaround-ed code path — the live regression test for Iterator compilation bug: reached an Unreachable point #769), skdb-wasm, and fast-checks all pass.

Merge plan

1. Bootstrap promoted from this branch's stage1 (make -C skiplang/compiler promote) and submitted as a separate PR — the normal process. This must land before this PR merges: the post-merge :latest images are built STAGE=0 from main's bootstrap, so an unpromoted bootstrap would ship a toolchain without the fix and break skdb tests on main.
2. After the bootstrap PR merges: rebuild and push the :latest images from main.
3. Drop the temporary ci: use fix-769 Docker images commit from this branch (CI then runs on :latest images, which by then contain the fix).
4. Merge.