Fix staged effect scope and prototype open rows

[MelbourneDeveloper]

Problem and result

Partial runtime handlers could hide outer static arms, generic effect instantiations could lose identity during discharge, and returned handler factories could reach codegen with incompatible erased callback types. This PR preserves operation scope and concrete callable types through validation, lowering and codegen. A written ![] rejects operations, runtime builtins and unproven callbacks in both flavors and module signatures.

The branch parses and checks a closed-program prototype of !e and ![Log | e]; formatting, generated docs and editor signatures preserve the written row. Runnable Osprey/Koka examples substitute two implementations for the same work. A staged example runs one open-row helper with a static or dynamic handler in both flavors. Static-only lowering has no effect-dispatch symbol; native, wasm32 and iOS corpus goldens pin its output. The installed VSIX test retries transient VS Code code-action cancellation only when the document version is unchanged.

Validation

• Current head 775cc99c: native corpus 217/217 under default, GC and ARC (zero ARC leaks); wasm32 corpus 151/151 with 66 pinned capability skips; iOS simulator corpus 136/136 with 81 pinned capability skips. Both new flavor cases print 41 on all three targets. API docs now show the declared row once with source-local names. The full local workspace suite passed on the preceding head; all 16 focused docs tests, the exact HTML integration check, formatting and strict workspace Clippy pass on this head. Hosted CI is restarting for this repair.
• The runnable handler, continuation, answer-transformation and open-row comparisons pass against real Koka, OCaml, Eff and Effekt where each supports the example.
• Local make ci stops at the repository's Deslop gate because its executable is unavailable on this machine; hosted CI runs the pinned action.

This PR remains draft. The open-row support is a closed-program prototype, not full row polymorphism: function types still lack independently quantified latent rows and scoped row unification. Plan 0016 still requires named instances and masking, owned/multi-shot continuations, deterministic finalization, wasm control-handler parity and inferred-effect tooling. The related structured-concurrency and arithmetic-effects plans also remain unimplemented. Green CI alone does not establish those contracts.