skillscript-runtime v0.26.6

Upgrade impact: none (additive; help-content + server-instructions only). No runtime behavior change.

The 0.26.5 error-handling topic taught a broken containment pattern: its flagship Rule 2 example put a per-leg (fallback:) on execute_skill legs and claimed it prevents a gather abort. It does not — and shipping it canonicalized the exact pattern that sank the 2026-07-05 morning brief (Perry c052581b, verified on the deployed 0.26.5 runtime).

• Corrected the mental model to two failure shapes → two tools. A MISSING VALUE (dispatch error, shell spawn-fail/timeout, empty result, unresolved ref) is contained by (fallback: "…") / a fallback filter. A RAISED THROW ($ json_parse on off-shape input; execute_skill whose child throws — e.g. the child's output template references an unset var) is NOT caught by (fallback:) — the throw propagates past the trailer and aborts the target. A raised throw is contained by an else: handler, or prevented by a structural guard (pre-bind degraded defaults + a contains/shape check before the risky op).
• Every prescription was gate-dogfooded on the runtime before shipping (the whole bug was a doc that confidently prescribed an unverified mechanism): confirmed (fallback:) catches neither an execute_skill child-throw nor a json_parse malformed-throw; confirmed else: catches both; confirmed the structural guard degrades cleanly. # OnError: was found inert (its fallbackSkillExecutor is never wired in the runtime) and is therefore flagged as not-a-reliable-throw-container rather than prescribed.
• Rule 2 now carries the corrected worked example (a throw-proof get-weather child), and the quickstart + server-instructions robustness lines were corrected to the same two-shapes framing.