Isolate SPAWN_AGENT body EXIT on -run-bc

[howlcipher]

Summary

Post-#102 swarm slice (Owner option: isolate): isolate VmExit from inside a SPAWN_AGENT task body on the independent -run-bc bytecode path so a bare exit is a child failure instead of escaping and aborting the parent run.

• Hypothesis verified: the child recover path in OpSpawnAgent re-panicked VmExit (case VmExit: panic(r)), which reached the top-level runner and became the process exit code (e.g. exit 7), with no child diagnostic and no parent continuation.
• Change: stop re-panicking VmExit. Emit one ErrOut diagnostic — [Swarm VM] Agent "Name" failed task: "desc": EXIT: exit N from spawn agent body — set the existing failed flag, skip that agent's completion line, restore parent stack / shared executed accounting as before, and let the parent/siblings continue (exit 0).
• Still fatal (unchanged): LIMIT_EXCEEDED (budget + depth), parent SPAWN_AGENT process gate, and non-string task TYPE_ERROR. Ordinary child *VMError isolation from Isolate nested SPAWN_AGENT child failures on -run-bc #101 and VmReturn isolation from Isolate SPAWN_AGENT body RETURN on -run-bc #102 are unchanged. CALL still recovers normal function returns inside a spawn body.
• Tests: added body exit isolated (parent print after exit 7) and nested exit isolated (inner exit; outer prints and completes). Existing nesting/isolation/RETURN cases remain.

Constraints honored

• No production -compile-bc → HFIR flip
• Tip-lock 4d74dbcf not retaken
• No DOM; Write flagged http_server bytecode and stop #90 stays Partial
• No new opcodes / no opcode metadata change
• No C1/C2/C4/harness scope creep

Verification

• gofmt -l internal/vm/vm.go internal/vm/spawn_agent_test.go (empty), git diff --check
• go vet ./internal/vm/ ./internal/bytecode/ ./examples/ .
• go build ./...
• go test ./internal/vm/ -count=1 -run 'TestVMTaskAndSpawnAgent|SpawnAgent'
• go test ./... -count=1 (full suite green)

Journal

docs/journals/2026-10-05_spawn_agent_exit_isolation.md

Base tip: 6502d615ef40025d3e4bdc9240a4cf54927fc6af (#108)

[howlcipher]

Motoko COMMENT on head 47a461abdee36a458382bd351aa86af1aebf1607 (base 6502d615).

Verdict: merge-ready once CI build is green. Diff matches the claimed leftover after #102: stop re-panicking VmExit from SPAWN_AGENT child recovery so parent/siblings continue.

Checked

• OpSpawnAgent recover switch e := r.(type): case VmExit now emits ErrOut diagnostic with e.code (field exists on VmExit), sets existing failed, skips completion line — same shape as VmReturn.
• LIMIT_EXCEEDED still re-panics; ordinary *VMError / default isolation unchanged; stack truncate + executed accounting still before the switch.
• Tests: body exit isolated and nested exit isolated parallel the RETURN cases (exit 7, parent continues, exit 0, no RuntimeFailure, no completion for failed child).
• Scope: one-line VM change + tests + journal/changelog. Hard nos intact: no prod -compile-bc / HFIR flip, tip-lock untouched, no DOM, #90 Partial, no C1/C2/C4/harness.

Non-blocker: older change_log.md RETURN bullet still says “VmExit and limits still propagate”; the new EXIT bullet supersedes the VmExit half. Harmless historical wording — optional one-line tidy later.

Draft left draft. Okabe: undraft + squash-merge when CI is green.