[P1] Uncaught errors are diagnostics rather than terminal failures #10

Reproduced during the 2026-09-04 language review at a20753c6cff96ee86758744e227a04460234a832. Priority: P1.

use mod.std
std.assert(false)
42

The optimized runtime reports completed=true, returns 42, and includes an assertion diagnostic. An unknown variable or calling nil followed by 42 behaves similarly. Errors inside a returned array can also leave the outer evaluation marked successful even though JSON conversion subsequently rejects the embedded error value.

This is consequential at an effect boundary:

use mod.std
use mod.tool
std.assert(false)
tool.call("text.echo", "called after failed assertion")

With the CLI's explicit --allow-echo grant, the tool audit records Allowed and the CLI exits zero. A separate two-step workflow reproduced the larger consequence: step guard contains std.assert(false); 42; step effect calls its separately granted echo tool. The CLI reports both steps succeeded, the workflow completed, and an allowed tool event. Thus this is not just misleading result formatting: the next step's authority becomes active after a failed guard.

In vm.rs, handle_errors unwinds to a try when one exists, but otherwise only drains diagnostics. Evaluation::succeeded checks the final value alone. The workflow relies on that predicate before handling output and advancing; see splash-workflow.

Terminate the current evaluation at the first uncaught runtime error. Represent completion, suspension, and failure explicitly. Changing only succeeded() to inspect diagnostics is insufficient: an effect may already have executed by the time that predicate is checked. Preserve the existing distinction between recoverable try/catch errors and uncatchable resource stops. Regression tests should assert that no later effect occurs, no later step activates, and a caught error can still recover deliberately.

Acceptance: implement the described correction, publish the resulting language contract, and add focused regression coverage including applicable capability, streaming-parser, and resource-limit cases.

Fix and executable regressions: PR #19. Workspace tests, explicit vendor suites, formatting, strict Clippy, and the Linux ARM64/RISC-V and Windows GNU compile checks pass locally. All GitHub Actions checks pass on f1e775f, including Linux tests, fuzzing, AddressSanitizer, and Miri. This issue remains open until the fix is merged.